Shades of Smiles: Creating Variants of Smiles from Neutral Images of Real Individuals - Method and Validation

Shades of Smiles: Creating Variants of Smiles from Neutral Images of Real Individuals - Method and Validation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Shades of Smiles: Creating Variants of Smiles from Neutral Images of Real Individuals - Method and Validation Jin Gao, Werner Sommer, Rasha Abdel Rahman, Wei-Jun Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6908842/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 19 Mar, 2026 Read the published version in Psychological Research → Version 1 posted 9 You are reading this latest preprint version Abstract Despite their crucial role in emotion research, facial expression databases primarily contain stimuli of basic emotions, rather than subtle, socially nuanced ones. Here, we introduce a rigorous yet easily applicable pipeline for synthesizing fine-grained emotional expressions—specifically, reward, affiliative, and dominance smiles—from neutral photographs, guided by Facial Action Coding System (FACS) principles and the perspective that expressions serve as social signals. From neutral images of 90 individuals (45 female, 45 male), we generated five expressions each (including neutral and disgust), and mirrored each image to address hemiface biases. In an online validation study, 13 participants rated each image on emotional content, arousal, and plausibility, yielding 26 ratings per expression and model. Rating results show that (1) reward smiles were most reliably recognized, while affiliative and dominance smiles tended to be confused with related expressions and that (2) arousal and plausibility ratings varied systematically across expression types and model gender. In summary, the suggested expression generation pipeline and the validated stimuli offer a robust method and a scalable dataset for research on fine-grained emotion recognition and emotional communication. facial expressions emotional stimuli emotion recognition FACSGen 2.0 social smiles Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Facial expressions are a valuable research tool for the study of emotions, as they convey a large range of affects states, including feelings, intentions, and desires (Horstmann, 2003). They can be precisely described by observable facial surface changes driven by specific patterns of muscle activity, referred to as Action Units (Ekman & Rosenberg, 1997; Klingner & Guntinas-Lichius, 2023). Because of these attributes, researchers have collected various facial expression databases providing standardized stimuli. However, acquiring images of non-standard emotional expressions across a sufficiently diverse pool of expressors remains challenging. In this study, we suggest a solution based on the artificial creation of custom-defined expressions from photographs of individuals displaying neutral expressions. Unless cartoon faces are used, most realistic facial emotion stimuli have been sourced from real human expressors (Ekman, 1976; Olszanowski et al., 2015; Tracy et al., 2009). For a review, see Dawel et al., (2022). However, integrating large datasets with high-quality expression stimuli presents significant challenges. For instance, AffectNet—a database of facial affect compiled from the internet using numerous emotion-related search queries—comprises over one million images with faces and facial landmark points (Mollahosseini et al., 2019). Nonetheless, these images are typically captured under uncontrolled conditions with considerable variability in lighting, pose, resolution, and background, and only a fraction of images depict a given emotion accurately. Specifically, only about 48% of images queried for "happy" expressions were annotated as happiness; percentages for other emotions were even lower (below 20%; see Mollahosseini et al., 2019, Table 4). In contrast, databases like the MPI Facial Expression Database (Kaulard et al., 2012) and the McGill Face Database (Schmidtmann et al., 2020) provide highly controlled stimuli obtained from trained models. However, training these expressors to reliably produce specific Action Units (AUs) or nuanced expressions—such as shame or composite expressions like a sad smile—requires considerable time and effort, and not all posers achieve consistent results (Mehu et al., 2012; Tottenham et al., 2009). For instance, the MPI Facial Expression Database includes an impressive array of 55 natural emotional and conversational expressions elicited through a method-acting protocol, in which participants were provided with specific everyday scenarios designed to evoke natural and authentic expressions; however, it comprises only 19 German models (Kaulard et al., 2012) Similarly, the McGill Face Database features 93 expressions of mental states but is based on merely two professional expressors (Schmidtmann et al., 2020). Consequently, for specialized research or specific applications, even these extensive databases are limited. Advances in generative adversarial networks (GANs) and related AI-driven methods have substantially improved the generation of high-quality facial stimuli (Kammoun et al., 2022). For instance, Deepfake technologies, based on GAN architectures, can produce highly realistic face images that are often indistinguishable from real photographs (Tolosana et al., 2020). Similarly, GANimation (Pumarola et al., 2020) offers the ability to manipulate facial Action Units (AUs) to generate expressions in a controllable manner. However, while both methods excel when the input closely matches the training data distribution, they tend to falter when applied to faces with distinctive structures or when generating subtle, nuanced emotional expressions. In particular, Deepfake technologies, although highly convincing in overall face appearance, often lack precise control over specific emotional expressions. Likewise, outputs from GANimation can appear distorted or unrecognizable when dealing with complex or unfamiliar facial features. These limitations, rooted in training biases, restrict the applicability of these methods in research contexts that require a large set of subtle, systematically varied emotional expressions across different individuals. Facial modelling and animation techniques from computer vision and graphics provide another solution for generating extensive sets of emotional facial expressions. Unlike physics-based simulations that rely on complex geometric and muscular representations (Behrouzi et al., 2025; Rai et al., 2024), the FaceGen software suite—particularly FACSGen—offers an accessible and intuitive platform for AU manipulation. FACSGen 2.0 integrates advanced algorithms with a user-friendly interface to map photorealistic textures onto virtual faces, enabling the precise generation of both static and dynamic expressions. Krumhuber et al., (2012) demonstrated the utility of this software in producing AU-defined facial expressions, including canonical emotions (e.g., happiness, anger, fear) and less common social emotions (e.g., embarrassment, contempt), across various genders and ethnicities using grayscale faces. Building on the work of Krumhuber et al. (2012), the present study focuses on three specific types of smiles—reward, affiliative, and dominance—that commonly function as social signals in interpersonal contexts (Hareli & Hess, 2012). These smiles are defined by distinct Action Unit (AU) configurations (Martin et al., 2017; Niedenthal et al., 2010). Reward smiles are displayed to express happiness, reward others or to communicate positive experiences or intentions; they involve the activation of the Lip Corner Puller (AU12), Inner-Outer Brow Raiser (AU1-2), Sharp Lip Puller (AU13), and Dimpler (AU14) (Rychlowska et al., 2017). Affiliative smiles are shown to signal appeasement and create and maintain social bonds and also involve AU12 but are distinguished from Reward smiles by activating the Lip Pressor (AU24) and Dimpler (AU14). Dominance smiles are displayed to negotiate status within and across social hierarchies; they also feature AU12 but asymmetrically, involving either the right or left Lip Corner Puller (AU12R or AU12L). They are also characterized by the Upper Lid Raiser (AU5), Nose Wrinkler (AU9), Cheek Raiser (AU6), and Upper Lip Raiser (AU10). To allow for manipulation checks and to differentiate positive from at least one negative expression, we also included disgust expressions. Disgust, a negative expression, shares the Nose Wrinkler (AU9) with the dominance smile (Pochedly et al., 2012). Furthermore, neutral expressions were derived from the model faces with the same pipeline but without AU activation, allowing these expressions to serve as experimental baseline. While Krumhuber et al., (2012) pioneered AU-based facial expression generation using FACSGen 2.0 , the present study addresses key limitations in their study and extends it to broader applications. First, whereas Krumhuber et al., (2012) used a small number of expressors (at most four expressors per experimental condition), our stimuli were derived from a much broader range of real front-facing neutral portraits sourced from multiple databases. These images were converted into standardized head expressors while preserving identity-specific features. Although this study focused on a single racial group to ensure experimental control, our method is scalable to more diverse populations. Second, although Krumhuber et al. (2012) included a few less-common social expressions (e.g., embarrassment, contempt), their work did not provide a comprehensive pipeline for generating and validating nuanced, socially meaningful variants of expressions. In contrast, our integrated pipeline systematically produces and validates three subtle smile types—reward, affiliative, and dominance—each capturing distinct social signals such as appeasement versus assertiveness. Third, building on the validation approach used by Krumhuber et al., (2012), we evaluated recognition accuracy, perceived arousal, and plausibility. Crucially, we introduced the arcsine-transformed Unbiased Hit Rate (UHR), a response-bias-resistant metric, and incorporated it into a composite index combining plausibility ratings, allowing for a more comprehensive assessment of model performance. Finally, to support reproducibility, we provide open-source validation code and share all stimuli derived from our in-house images. We hypothesised that 1) participants would reliably recognize and accurately identify the distinct emotional expressions depicted in the synthesized facial images, that is, "reward smiles," "affiliative smiles," "dominance smiles," "disgust," and "neutral expressions." The accuracy rates for identifying these expressions were expected to be significantly better than chance, indicating clear and distinguishable emotional contents. 2) Smiles and disgust were expected to be associated with higher perceived arousal levels than neutral expressions. 3) The synthesized images were expected to be perceived as highly plausible, with participants judging the images as being derived from photographs of real humans. Methods Construction Stage To create the emotional expression stimuli, we followed a structured series of steps as illustrated in Fig. 1. Step 1: selection and preparation of database sources. As original images we selected 90 high-resolution front-facing photographs of Caucasian individuals with neutral facial expressions from several sources (DeBruine & Jones, 2017; Ebner et al., 2010; Ma et al., 2015; Rellecke et al., 2012; Strohminger et al., 2015). Stimuli comprised 45 male and 45 female expressors, with mean ages of 25.7 and 26.3 years, respectively. While there were slight variations in parameters across picture sources (e.g., the colour of backgrounds and the size of the pictures), they were sufficiently similar to provide a consistent basis for the generation of head models since they were, taken under similar shooting conditions and ensured clear visibility of individual facial features. Besides, we selected only models without beards or non-face features, such as, scars, tattoos or piercings, jewellery, glasses, makeup or any clothing that covered the neck. Step 2: generation of head models. We utilized the FaceGen Modeller, a part of the FaceGen software suite (Roesch et al., 2011), allowing to generate a virtually infinite number of 3D facial identities as head model. The selected photographs were imported into FaceGen Modeller software. The software extracted facial geometry and skin texture information from these images, constructing detailed 3D head models with realistic skin textures and individual-specific features. Specifically, photorealistic skin texture was mapped onto the face. Different texture layers (i.e., diffuse colour, ambient occlusion, and gloss and normal maps) were also included, which are combined during the rendering stage to achieve the final appearance (Krumhuber et al., 2012). Step 3: generating and adjusting facial expressions based on action units (AUs). FACSGen 2.0 software (Krumhuber et al., 2012) allows detailed adjustment and combining various facial expressions based on Action Units (AUs), enabling researchers to simulate a range of emotional expressions without altering the identity of the expressors. We began the process by selecting specific combinations of AUs that correspond to the targeted facial expressions "reward smile," "affiliative smile," "dominance smile," and "disgust." The AU combinations for each expression (see Table 1) were selected based on the smiling prototypes described by Martin et al. (2017) and the disgust prototype (Pochedly et al. 2012). Additionally, since all original portraits displayed neutral expressions, the 'neutral expression' of the 3D models was directly generated from the imported photographs without any manipulation of AUs. Table 1 Expression AU combinations reward smile AU1 + AU2 + AU6 + AU7 + AU12 + AU13 + AU14 + AU20 + AU26 + ahh + i + k affiliative smile AU6 + AU7 + AU12 + AU13 + AU14 + AU20 + AU24 dominance smile AU2 + AU4 + (AU1+2+4) + AU5 + AU6 + AU7 + AU9 + AU10 + AU12+ AU14_Unilateral + AU25 + AU17 disgust AU4 + (AU1+4) + AU6 + AU7 + AU9+ AU10 + AU11+ AU16 + AU15 + AU17 + AU20 + AU24 Targeted Facial Expressions for Each Emotion and Their Corresponding Action Units (AUs) Notes. AUs in boldface are considered essential for inclusion in the final portrayal of a given expression; AUs in normal typeface can occur at any level of intensity. AUs named “ahh”, “I”, “k”, and those in parentheses are pre-defined combinations of AUs in FACSGen 2.0, which can be used to fine-tune the expressions to better approximate their prototypes. In FACSGen 2.0, each AU is represented by a slider controlling the magnitude of the morph target, which gradually transforms the original neutral expression (0%) into the fully activated target expression (100%). We used the slider-based controls to adjust the intensity of the AUs specified for the target emotion, fine-tuning them to closely match the prototype of each specific expression. This was necessary because setting the AU activity at a preset level for all faces may induce artificial looking results. In adjusting the AU strengths we followed the principle of keeping the expressions as natural as possible without unnatural facial distortions, such as teeth-lip intersections or excessive muscle activation. For AU12—activated in all three types of smiles and playing a crucial role in defining the characteristic look of each smile—the intensity was uniformly set to "80%" for all three smiles (without manual fine-tuning). It is important to note that FACSGen does not provide asymmetric AU12 activation, which is considered to be crucial for the dominance smile (Martin et al., 2017; Niedenthal et al., 2010). To achieve this effect, we employed a compensatory method by combining "dragging one side of the mouth corner upward in FaceGen Modeller on the face picture" and increasing the intensity of AU14_Unilateral in FACSGen to approximate the asymmetric AU12 effect. Step 4: standardization of display parameters and generation of static images. The original hairstyle and hair colour of the individuals in the photographs cannot be retained during the 3D model generation process because the FACSGen 2.0 software does not support importing or replicating external features such as hairstyle and hair colour. To further standardize the images, we ensured uniformity of hairstyle and hair colour of the generated head models. Note, there are many hair styles and colours available in FACSGen that could be used. For present purposes, all individuals of a given sex were assigned the same sex-typical hairstyle and all models, regardless of sex, were given a uniformly brunette hair colour. Additionally, the eye colours of the models were set according to the frequencies typically observed in Caucasian populations, that is, 40% brown, 30% blue, 20% green and 8% grey (Sulem et al., 2007). Display parameters, including frontal face orientation, taupe background, and lighting conditions, were uniformly set for all 3D models. To simplify evaluation and communication, we exported 2D front view projections of the models at a resolution of 800 × 1200 pixels with 32-bit depth. The 2D projections captured the key features of the 3D models in a single, clear image, making it straightforward to identify essential characteristics and potential issues. These images were then used for validation. Validation Stage Participants. As detailed in our pre‐registration 1 ，we conducted an a priori power analysis in G*Power 3.1 for our mixed‐design ANOVA (within‐between interaction) with the following parameters: medium effect size ( f = 0.25), α = .01, power (1 − β ) = .90, total groups = 10 (5 lists × 2 versions), repeated measures = 5 (emotion categories), assumed correlation among repeated measures = .50, and non-sphericity correction ( ε ) = 1. This yielded a minimum of 80 participants (8 per group). To enhance the precision of item‐level variance and confidence intervals, we increased to N = 130 raters (13 per group), exceeding the a priori requirement and ensuring robust power for all planned contrasts. Accordingly, via Prolific (https://www.prolific.com/), we initially recruited N = 135 participants (paid £7.62/hr). All were right‐handed adults (18–34 years) with normal or corrected vision, no reported physical or mental health issues, and self‐identified as White to preclude other‐race effects given our Caucasian models (Crookes et al., 2015; Zhao et al., 2024). The study adhered to the Declaration of Helsinki and received ethics approval from the Department of Psychology at Humboldt-Universität zu Berlin. Participants gave informed consent, including agreement to non-dissemination of images. After excluding one rater who reported “Other” gender and four with below‐chance accuracy, the final sample comprised 130 participants: 70 women ( M = 25.73 years, SD = 4.30) and 60 men ( M = 25.68 years, SD = 5.02). Experimental stimuli. The images used had been synthesized from portraits of 90 expressors (45 males and 45 females) with neutral expression (according to the normative data of the sources). For each portrait, five images with different expressions were created: "reward smile," "affiliative smile," "dominance smile," "disgust," and "neutral expression." For clarity in describing the assignment scheme, the synthesized emotional expressions are referred to as if they had been displayed by the expressors themselves, although they were generated from their neutral portraits. To control for potential hemiface effects—a phenomenon referring to asymmetry biases in human emotion recognition, where there is preferential looking at the left hemiface when actively exploring face images (Borod et al., 1997; Busin et al., 2018), we created two mirrored versions of each facial image, labelled as L-version and R-version, respectively. To minimize rater fatigue but ensure comprehensive coverage of all images, we divided the 90 model identities into five model groups of 18 (9 males and 9 females) each. A Latin Square design assigned each model in a group to display only one of five expressions (reward, affiliative, dominance, disgust, neutral) in the original or mirrored version, resulting in 5 different lists of images. For each of these five list a second list was created with mirrored versions of the images, yielding a total of 10 image lists. For example, in List 1, expressors in Groups 1–5 displayed reward, neutral, disgust, dominance, and affiliative, respectively; in List 2, the same expressors displayed neutral, disgust, dominance, affiliative, and reward, respectively. This rotation continued through all ten lists; hence within a list each model had presented only one expression but across lists each model had shown all expressions in original and mirrored versions. Since a given rater received only one of the 10 lists, we avoided crosstalk across different expressions in a given model within the same rater. Each of the ten lists was presented to a distinct group of 13 participants, such that every image was rated by 13 different individuals. Procedure. The ratings were conducted online using the SoSci Survey platform. Raters were informed that they were taking part in a validation study of emotional pictures of human faces. Before the main task, participants were instructed to adjust their viewing distance to their screens at home to standardize the visual angle of the facial images; the horizontal width of the face image should roughly correspond to the combined width across index, middle, and ring fingers with arms outstretched. This arrangement yields a visual angle of approximately 4° horizontally, as used in many experiments. The instructions emphasized that there are no 'right' or 'wrong' answers and participants were encouraged to base their responses on their spontaneous and subjective opinions but to pay close attention to the (sometimes) subtle differences between the facial expressions. During the practice phase, participants were familiarized with the six expression categories used in the task: reward smile, affiliative smile, dominance smile, neutral, disgust, and other. These categories were operationally defined to support consistent judgment, based on brief descriptions and everyday scenarios (e.g., reward smile: "expressing positive internal happiness, for example, after receiving good news"; affiliative smile: "inviting and maintaining mutually positive and beneficial social bonds, typically expressed when greeting a stranger"; dominance smile: "instantiating, stabilizing, or maintaining dominance in a social relation, for example, when disapproving the ideas or actions of a person viewed as intellectually or physically inferior"; disgust: "rejection or revulsion of something or someone potentially contagious, offensive, distasteful, or unpleasant, for example, seeing someone vomiting"; neutral expression: "a non-expressive posture of the face"; other expression: "not belonging to one of the five previous categories"). As illustrated in Fig. 2, for each image, three questions were posed in order: 1) "Which facial expression does this person primarily show?" (emotional content): Here, the participants should choose one of six options: reward smile, affiliative smile, dominance smile, neutral, disgust, and other. The ordering of the expression labels was kept constant within each rater but was counterbalanced across participants. 2) "How strong is this person's level of arousal?". For these arousal ratings we used the Self-Assessment Manikin (SAM; P. Lang, 1980), a pictorial technique with graphic figures representing different levels of arousal, ranging from a completely relaxed, unstimulated, and calm state (depicted by a manikin with eyes closed) to a highly awake, stimulated, and excited state. 3) "How likely is it that this face is based on a real human photograph?" (plausibility). Raters selected one of seven probability levels, ranging from "Impossible" to "Certain". After each answer, participants clicked the "Next" button to proceed to the subsequent question. For Questions 2 and 3, participants were encouraged to utilize the full range of the provided rating scales. After completing several practice tasks, participants proceeded to the main experiment. Each image was presented along with the rating questions on the same screen, allowing participants to respond immediately after viewing the image. There was no time limit for completing the ratings, and participants were encouraged to take breaks whenever needed at their own discretion. Throughout the experiment, five attention check questions were randomly interspersed, as suggested by Prolific, to ensure participants remained engaged and responded attentively. These questions explicitly requested to complete a task in a specific way, for example “The colour test you are about to take part is very simple, when asked for your favourite colour you must select 'Red'”. Data Analyses and Results In the following we report results for the emotional category ratings as simple hit rates (SHR), unbiased hit rates (UHR), and as confusion matrix. Then we report the results for the arousal and plausibility ratings, followed by an analysis of interrater reliability. All analyses were conducted using R Statistical Software (v4.3.3; R Core Team 2024). Simple Hit Rate SHR is a classic accuracy index, which is simple and intuitive (Bänziger et al., 2012). In the current study, SHR were mean percentages of correct responses indicating how often the chosen emotion agrees with the experimenter-intended emotional expression. SHR was calculated separately for each emotion category across all participants and expressors. Descriptive statistics for SHR across expression categories are summarized in Table 2. The SHR was highest for neutral ( M = 92.6%, SD = 11.3), reward ( M = 92.1%, SD = 14.7), and disgust expressions ( M = 91.5%, SD = 13.5), and substantially lower for affiliative ( M = 69.2%, SD = 29.5) and dominance smiles ( M = 66.0%, SD = 28.8). Given the six-alternative forced-choice format, the theoretical chance level was 16.7%. All mean SHR values were markedly above chance, suggesting that participants were able to recognize the expressions at rates well beyond guessing. Table 2 Expression Type SHR M (%) SHR SD UHR M UHR SD Chance UHR M Chance UHR SD Neutral 92.6 11.300 0.880 0.160 0.044 0.012 affiliative 69.2 29.500 0.594 0.305 0.038 0.016 Disgust 91.5 13.500 0.892 0.166 0.042 0.009 dominance 66.0 28.800 0.648 0.309 0.031 0.014 Reward 92.1 14.700 0.798 0.219 0.049 0.016 Mean (M) and Standard Deviation (SD) of Recognition Accuracy Measures by Expression Type Note . SHR = Simple Hit Rate (percentage of correct responses); UHR = Unbiased Hit Rate; Chance UHR is computed based on the marginal frequencies of the confusion matrix for each expression type (Wagner, 1993). SHR data were analysed using a binomial (logit) generalised linear mixed model (GLMM). This model was chosen because SHR values are based on trial-level binary outcomes (correct/incorrect), which naturally follow a binomial distribution. Attempts to fit a linear mixed-effects model (LMM)—even after arcsine transformation—violated assumptions of normality and homoscedasticity. A GLMM thus provided a more appropriate framework for analysing the binary nature of the data while accounting for random variability at the rater level. In this model, expression category, group, rater gender, model gender, and version were specified as fixed effects, and rater (case) was included as a random intercept. The model showed good overall fit (AIC = 8854.1, BIC = 8986.7). Relative to neutral expressions, both affiliative ( β = –1.95, SE = 0.10, z = –20.27, p < .001) and dominance smiles ( β = –2.12, SE = 0.10, z = –22.16, p < .001) were associated with significantly lower probabilities of a correct response. Disgust ( β = –0.16, SE = 0.11, z = –1.41, p = .16) and reward smiles ( β = –0.07, SE = 0.11, z = –0.58, p = .56) did not differ significantly from neutral. No significant effects were found for group, rater gender, face gender, or version (all p > .10). Model diagnostics using the DHARMa package indicated no signs of overdispersion ( p = .984) or zero-inflation ( p = .938), supporting the adequacy of model fit. Tukey-adjusted post-hoc comparisons showed that SHR for affiliative and dominance smiles were significantly lower than for neutral, reward, and disgust expressions (all p < .001). Additionally, affiliative expressions yielded significantly higher SHR than dominance expressions ( p = .075, marginal). Unbiased Hit Rate Accuracy estimates are affected by two kinds of biases - the relative number of items for each emotion category presented (presentation bias) and the relative utilization of different emotion categories by the participants (response bias) (Bänziger et al., 2012). UHR is an accuracy index, which considers both biases by calculating “the joint probability that a stimulus is correctly identified (given that it is presented) and that a response is correctly used (given that it is used)” (Wagner, 1993). Here we calculated the UHR following Bijsterbosch et al. (2021) in two steps: First, a choice matrix was created with chosen emotion expressions as columns and intended emotional expressions as rows. Second, the number of ratings in each cell was squared and divided by the product of the marginal values of the corresponding row and column (Note: there were no missing data). Since UHR is a proportion, the values were arcsine-transformed to correct for skewed variances of decimal fractions obtained from count (Bishara & Hittner, 2012; Goeleven et al., 2008). The transformed scores can range from 0 to 1.57, the arcsine of 1, representing perfect identification. Descriptive statistics for arcsine-transformed UHRs followed a similar pattern with SHRs (summarized in Table 2). Neutral, reward, and disgust expressions achieved high UHRs (e.g., arcsine M = 1.42, 1.37, and 1.39 respectively), whereas affiliative and dominance yielded lower performance (arcsine M = 1.00 and 1.03). Similarly, unbiased hit rates (UHR) exceeded chance levels across all expressions, further confirming the recognizability of the synthesized emotional stimuli. To compare the UHRs across emotion categories, we applied a linear mixed effects model (LMM) on arcsine-transformed unbiased UHRs, treating expression category, group, rater gender, model gender, and mirror version as fixed effects, and including a random intercept for each rater (CASE). The model converged with a REML criterion of 732.7. Relative to the reference category (neutral), expression type showed significant effects. Specifically, compared to neutral, the affiliative condition yielded a significantly lower UHR ( β = -0.425, SE = 0.026, t = -16.50, p < 0.001), as did the dominance condition ( β = -0.394, SE = 0.026, t = -15.26, p < 0.001). In addition, reward exhibited a small yet significant effect ( β = -0.052, SE = 0.026, t = -2.02, p = 0.043), whereas disgust did not differ from neutral ( β = 0.012, SE = 0.026, t = 0.48, p = 0.630). Other factors (face gender, rater gender, version, and group) did not significantly affect UHR (all p > 0.10). Post-hoc pairwise comparisons (Tukey-adjusted) revealed that neutral differed significantly from both affiliative ( estimate = 0.425, t = 16.50, p < 0.0001) and dominance ( estimate = 0.394, t = 15.26, p < 0.0001), whereas differences between neutral and either reward ( estimate = 0.052, t = 2.02, p = 0.2553) or disgust ( estimate = -0.012, t = -0.48, p = 0.9890) were not significant. Additionally, affiliative differed significantly from both disgust ( estimate = -0.437, t = -16.98, p < 0.0001) and reward ( estimate = -0.373, t = -14.47, p < 0.0001), while the difference between affiliative and dominance was non-significant ( p = 0.7413). Further comparisons (e.g., disgust vs. dominance: estimate = 0.406, t = 15.74, p < 0.0001; dominance vs. reward: estimate = -0.341, t = -13.24, p < 0.0001) underscore the robust impact of expression type on UHR. Overall, our findings on UHR aligns with SHR: neutral, disgust, and reward conditions yield higher recognition performance than affiliative and dominance, while group assignment, rater gender, model gender, and version have no systematic impact. Confusion Matrix According to Bänziger et al. (2012) a Confusion Matrix is an appropriate and complete way to provide all the relevant information about potential bias since different attempts to propose indices that are corrected for chance guessing and/or response bias (e.g., the UHR suggested by Wagner, 1993) generally mask the origin of the potential biases and do not allow to take these into account in interpreting the results. Thus, we also present the confusion matrix (see Fig. 3) representing the classification results of emotional facial expressions, illustrating the percentage of chosen emotional categories (e.g., reward, disgust, affiliative, dominance, neutral, other) for each intended emotional expression. A close look at the confusion matrix reveals several notable patterns in the participants' responses to different facial expressions. Faces with intended affiliative smiles were frequently confused with neutral expressions (7.5%) and reward smiles (17.7%), suggesting significant perceptual overlap between these expressions. Similarly, intended dominance expressions were misclassified as affiliative (10.9%) disgust (7.8%) or others (9.8%), highlighting some confusion in identifying dominance-related expressions. Reward expressions were correctly identified in most cases, though there was some misclassification as affiliative (5.9%). Arousal and plausibility In our study, the mean scores of the arousal ratings were 2.63 ( SD = 1.66) for neutral expressions, 4.50 ( SD = 1.60) for affiliative, 5.44 ( SD = 1.66) for disgust, 5.29 ( SD = 1.59) for dominance, and 6.72 ( SD = 1.39) for reward. Mean scores for plausibility ratings were 5.21 ( SD = 1.26) for neutral expressions, 4.66 ( SD = 1.39) for affiliative, 4.23 ( SD = 1.41) for disgust, 4.06 ( SD = 1.43) for dominance, and 5.02 ( SD = 1.33) for reward. To compare the arousal and plausibility scores across expression categories, linear mixed‐effects models (LMM) were fitted to these scores with fixed effects of expression category, group, rater gender, model gender, and image version, and a random intercept for CASE. Models were fitted using restricted maximum likelihood (REML). The intercept was estimated at 2.759 ( SE = 0.223, t (≈123.9) = 12.36, p < 0.001), representing the predicted arousal score for the reference level (neutral) of expression category. Compared with neutral expressions, the fixed effects estimates for expression category were as follows: affiliative: 1.865 ( SE = 0.0408, t = 45.73, p < 0.001); disgust: 2.808 ( SE = 0.0408, t = 68.88, p < 0.001); dominance: 2.661 ( SE = 0.0408, t = 65.26, p < 0.001); and reward: 4.088 ( SE = 0.0408, t = 100.26, p < 0.001). Tukey‐adjusted pairwise comparisons confirmed that ratings in the neutral condition were significantly lower than each other expression (all p -values < 0.001). In comparisons between non-neutral categories, all differences were significant ( p < 0.001), with the exception of the disgust–dominance contrast, which, though smaller, remained statistically significant ( p = 0.0028). Additionally, the effect of model gender was significant: images showing male faces received lower arousal ratings than those showing female faces (coefficient = –0.145, SE = 0.0365, p < 0.001); A similar LMM was applied to the plausibility scores. The intercept was estimated at 4.833 ( SE = 0.190, t (≈124) = 25.46, p < 0.001). Compared with neutral expressions, the fixed effects estimates were –0.553 ( SE = 0.035, t = –15.79, p < 0.001) for affiliative, –0.986 ( SE = 0.035, t = –28.15, p < 0.001) for disgust, –1.150 ( SE = 0.035, t = –32.84, p < 0.001) for dominance, and –0.197 ( SE = 0.035, t = –5.61, p < 0.001) for reward. Tukey‐adjusted comparisons indicated that all pairwise comparisons between expression categories were statistically significant ( p < 0.001). Moreover, the effect of model gender was significant, with images of male faces receiving higher plausibility ratings than female faces (coefficient = 0.248, SE = 0.0313, p < 0.001). Additionally, stimulus version (original vs. mirrored) significantly affected plausibility ratings, with R-versions (mirrored images) rated slightly but significantly less plausible than L-versions (original orientation) (coefficient = –0.169, SE = 0.044, p < 0.001). Overall, although the Shapiro–Wilk and Levene tests for the residuals were statistically significant—likely due to the large sample sizes—visual inspection of residual plots suggested that deviations from normality and homogeneity of variance were minor. Overall, both LMM models revealed statistically significant differences in ratings across expression category levels, and the potential effects of model gender in shaping perceptual evaluations of the images. Interrater reliability To assess the reliability of the ratings across participants for arousal, plausibility, and emotion categorization, we calculated Fleiss' Kappa (for categorical data) and Intraclass Correlation Coefficients (ICCs) (for continuous data). Fleiss’ Kappa estimates the agreement among multiple raters for nominal categories (Fleiss, 1971), while ICCs provide a measure of consistency in ratings across raters (Koo & Li, 2016; Shrout & Fleiss, 1979).A linear mixed-effects model with Group (i.e., each of the 10 rater lists) as a fixed effect was first used to examine potential between-group differences. Since no significant group effects were found for any of the rating types, we report mean Kappa and ICC values collapsed across all 10 groups. Fleiss' Kappa. The agreement between participants in categorizing emotional expressions was substantial, with an overall Fleiss’ Kappa of 0.65, indicating reliable identification of the intended expression categories. ICC. To objectively select the appropriate ICCs, the model fit (deviance information criterion) was calculated for all possible ICC models. The model with the best fit to our data was ICC 2, applying a two-way cross-classified multilevel model. The reliability of the judges' ratings for arousal and plausibility was assessed using ICC (2,1) and ICC (2, k ). ICC (2, 1) intraclass correlation coefficient, each image is measured by each rater with reliability estimated from a single measurement; ICC (2, k ) as before, but here reliability is calculated by imputing the average of the k raters’ measurements (Bijsterbosch et al., 2021). For arousal, the mean ICC (2,1) was 0.44, indicating moderate agreement between individual raters, while the ICC(2,k) was 0.97, reflecting excellent reliability at the group level. For plausibility, ICC (2,1) was lower at 0.14, indicating poor reliability across individuals, but the ICC(2,k) reached 0.88, suggesting good average agreement across raters (based on standard guidelines from Koo & Li, 2016). Ranking of Expressors In order to assess the relative quality of the expressions derived for the different individuals on which the expressions were based we ranked the performance of all expressors. At first we calculated the UHR and plausibility scores across all emotion categories per expressor. As described above, the UHR values had been arcsine-transformed to stabilize variance. A composite score was then derived by averaging the Z-scores of the arcsine-transformed UHRs and the plausibility scores. Our 90 expressors were ranked based on this composite score, grouped into ten rank-based categories, with gender distributions illustrated through pie charts (see Fig. 4). See Appendix for an overview of all individual expressors. Discussion Extending prior applications of FACSGen, the present study introduced a pipeline for synthesizing socially nuanced emotional expressions from neutral portraits of real individuals. We constructed a rigorously validated stimulus set of 900 images, with careful control of visual confounds. Recognition accuracy was high overall—particularly for reward and basic expressions—while affiliative and dominance smiles were more frequently confused with related categories, reflecting their subtle and socially overlapping nature. Arousal and plausibility ratings also varied systematically by expression type and model gender. These findings demonstrate not only the technical feasibility of AU-based expression synthesis but also its potential to advance research on subtle, socially meaningful facial expressions. Emotion recognition accuracy Recognition accuracy is a crucial measure of validity for emotional stimulus sets. In our validation study, we systematically evaluated how well emotional expressions synthesized from neutral images of real expressors—particularly subtle variants of smiles—can be identified. To this aim, we employed Simple Hit Rates (SHR) and Unbiased Hit Rates (UHR) to evaluate overall recognition accuracy, and visualized response tendencies using a confusion matrix illustrating perceptual overlaps among certain expressions. The mean SHR for neutral expressions was remarkably high (92.6%), surpassing previously reported hit rates, such as the 87% reported by Ebner et al., ( 2010 ). This likely reflects the fact that many of the neutral-expression portraits used here were sourced from validated datasets and had already undergone a selection process based on recognizability. The demographic profile of our participants—exclusively younger adults—may have also contributed, as younger individuals typically perform better in facial emotion recognition than older adults. Among emotional expressions, disgust demonstrated high recognition accuracy (91.5%), exceeding previously reported benchmarks such as the 68.6% accuracy reported by Krumhuber et al. ( 2012 ) during the initial validation of FACSGen 2.0-generated disgust expressions. The 'happiness' category used in previous studies closely aligns with the reward smiles in our study. The mean recognition accuracy for reward smiles in the present study (92.1%) was somewhat higher than the happiness recognition accuracy reported by Krumhuber et al., ( 2012 ) (88.46%). These findings highlight the efficacy of using carefully controlled AU configurations in FACSGen 2.0 to generate highly distinguishable emotional stimuli. Not surprisingly, recognition performance differed markedly from reward smiles for more subtle variants of smiles; SHR for affiliative and dominance smiles was at 69.2% and 66.0%, respectively. The confusion matrix revealed significant perceptual overlaps among these subtle expressions: affiliative smiles were frequently misclassified as reward smiles (17.7%) or neutral expressions (7.5%), while dominance smiles were commonly confused with affiliative smiles (10.9%), disgust (7.8%), or categorized as "other" (9.8%). These misclassification patterns suggest inherent perceptual similarities likely due to overlapping Action Unit (AU) configurations, aligning with previous findings by Rychlowska et al. ( 2017 ) indicating that subtle variations in smiles pose perceptual challenges. The observations from SHR and the confusion matrix were supported by the Unbiased Hit Rate analyses, which account for both presentation and response biases. Neutral expressions, disgust, and reward expressions consistently yielded higher recognition performance compared to affiliative and dominance smiles, though importantly, accuracy remained far above chance levels for all emotion categories employed here. In summary, our findings confirm the effectiveness of predefined AU configurations in producing distinguishable emotional expressions via FACSGen 2.0. At the same time, they emphasize specific perceptual challenges associated with subtle variants of smiles, reinforcing previous research and highlighting important areas for future methodological refinement. Arousal and plausibility Our analyses of arousal ratings revealed clear distinctions among expression categories that align with expectations. Consistent with previous research (e.g., Goeleven et al., 2008 ), neutral expressions elicited the lowest arousal ratings for the image ( M = 2.63), serving as an effective baseline for other emotional expressions. In contrast, disgust expressions produced substantially higher arousal ratings ( M = 5.44), reflecting their established role in signalling negative affective states (Rozin et al., 1994 ). Notably, our primary interest concerned the subtle variants of smiles. Reward smiles received the highest arousal ratings ( M = 6.72), which is consistent with their communicative function of conveying positive emotional intensity and social engagement (Sauter, 2010 ). Affiliative ( M = 4.50) and dominance smiles ( M = 5.29) elicited intermediate arousal levels, consistent with their more nuanced communicative roles. Affiliative smiles, aiming primarily at signalling warmth and cooperation, understandably evoke more moderate arousal. Dominance smiles, signalling social status or assertiveness, produced higher arousal than affiliative but lower than reward smiles, reflecting their social signalling function as documented in prior studies (Rychlowska et al., 2017 ). Plausibility ratings about how likely it was that the synthetic images were derived from pictures of real people were conducted to assess the apparent naturalness of the stimuli, which is very important for their usefulness in experimental studies (Diel et al., 2021 ). The plausibility ratings indicated distinctions among the expressions. Neutral expressions were perceived as most plausibly human-based ( M = 5.21), given their minimal modifications and frequent occurrence in natural interactions. Synthetic reward smiles ( M = 5.02) obtained high plausibility ratings comparable to neutral faces, consistent with findings that smiles reflecting genuine positive affect tend to be perceived as natural and authentic (Ekman & Friesen, 1982 ). In contrast, disgust ( M = 4.23) and dominance smiles ( M = 4.06) received lower plausibility scores, possibly due to their reliance on less frequently encountered Action Units (e.g., AU9 and AU10), making them less common in everyday facial displays and thus perceived as less genuine or natural (Dyck et al., 2008 ). The plausibility of affiliative smiles ( M = 4.66) fell in-between, suggesting that subtler smile expressions, while socially meaningful, might challenge observers’ perceptions of naturalness due to their ambiguity or less frequent occurrence (Rychlowska et al., 2017 ). Beyond expression type, we also observed systematic differences based on the model's gender, despite identical AU parameters across expressors. Specifically, female expressors elicited significantly higher arousal ratings compared to male expressors, aligning with previous research suggesting observers perceive female emotional expressions as more affectively intense or expressive (Hess et al., 2000 ; Plant et al., 2000 ). Conversely, male expressors received higher plausibility ratings, indicating their emotional expressions were perceived as more genuine or believable, though relatively unnatural hairstyles assigned to female expressors might partially explain this finding. Future studies could address this by concealing external facial features with oval masks or using more realistic hairstyles. Alternatively, societal stereotypes or expectations may also contribute to this result, as emotional displays by men could be seen as more intentional or credible, whereas heightened expressivity in women might sometimes be interpreted as exaggerated or performative (Kring & Gordon, 1998 ). Overall, these findings confirm that our synthesis method successfully generates expressions that vary meaningfully along both arousal and plausibility dimensions. The clear differentiation between reward, affiliative, and dominance smiles highlights the perceptual relevance of subtle AU variations in conveying affective meaning. Moreover, the observed gender differences in perceived intensity and naturalness, despite identical AU settings, suggest that visual appearance and social expectations interact in shaping emotion perception. Together, these results underscore the importance of considering both low-level visual cues and higher-order social factors in the design and validation of emotional facial stimuli. Interrater reliability Our analysis of interrater reliability indicates a generally robust consistency among participants. The overall Fleiss' Kappa value of 0.65 for emotion categorization demonstrates substantial agreement (Landis & Koch, 1977 ), suggesting that participants were consistent in identifying the intended expressions provided in the current study. This level of agreement supports the validity of our categorization process; for arousal ratings, the single-rater reliability, as measured by ICC (2,1), was moderate at 0.44, while the average reliability across participants, ICC (2, k ), reached an excellent level of 0.97. This disparity highlights that, although individual ratings might vary, the aggregate mean of multiple participants yields highly reliable arousal estimates; In the case of plausibility ratings, a similar pattern emerged: the single-rater ICC (2,1) was low (0.14), indicating considerable variability at the individual level. However, the aggregate reliability (ICC (2, k )) was strong at 0.88. These results underscore the importance of aggregating ratings in studies of this kind, as group averages effectively mitigate individual differences and enhance overall reliability. Collectively, these reliability indices confirm that our methodological approach yields consistent ratings across multiple evaluators, thereby reinforcing the robustness of our experimental design and the validity of our subsequent analyses. Ranking of Expressors Additionally, to further evaluate and compare the overall quality of synthesized stimuli, we developed a composite ranking score for each expressor by integrating standardized unbiased hit rates (UHR) and plausibility ratings. This composite measure enabled us to systematically identify which expressors performed optimally across multiple evaluation criteria, providing practical guidance for researchers selecting the most effective emotional stimuli for experimental purposes. The ranking also highlighted potential differences related to model gender distribution with images derived from male expressors scoring higher than female expressors, which could inform future studies aimed at balancing stimuli to mitigate biases or investigate gender-specific effects in emotion perception. We speculate that the generally smoother appearance of female faces makes it harder to effectively apply the AU manipulation by the software employed. Perspectives and Limitations The current procedure offers researchers considerable flexibility, allowing image selection based on various parameters—such as model gender, hemiface, expression category, recognition accuracy, arousal, and plausibility—to meet diverse research needs. Naturally, the pipeline can be extended to generate other or compound emotions, provided that the intended emotional expressions can be precisely defined in terms of Action Units (AUs). Moreover, our open-source code 2 base enables researchers to independently validate their own stimuli. Another promising avenue for future research is the inclusion of dynamic facial expressions, which can be readily created using FACSGen 2.0. Prior research has shown that dynamic expressions enhance affect recognition, increase perceived authenticity, and aid in distinguishing genuine from posed expression (Krumhuber et al., 2013 ). Additionally, expanding the database to include faces from more diverse racial backgrounds is both feasible and beneficial. Because our method operates on neutral base portraits, which are readily available or easily obtained, extending this framework requires no specialized acting skills or training on the part of expressors. Several limitations of the present work should be acknowledged. First, because our method is based on neutral portraits sourced from different existing databases, any stimuli originating from external databases are available only by contacting the respective database owners due to redistribution restrictions (see Appendix). However, since the validation analysis code is open-source, researchers can easily apply the synthesis pipeline and validation procedures to their own image sources. Additionally, experimental stimuli generated from our in-house database are directly available upon request, facilitating adaptation or extension to meet diverse research objectives. Second, although mirrored (R-version) and original (L-version) images did not differ in recognition accuracy or arousal, plausibility ratings were slightly lower for mirrored versions, especially for dominance and reward smiles. To objectively acknowledge and accurately report this observed difference, we decided to treat the two versions separately instead of merging the mirrored and original versions as initially intended. Consequently, the number of raters per expression per individual was reduced from 26 to 13. While this sample size per image aligns with typical standards in facial-expression validation studies (Delicato, 2020 ), it nonetheless raises potential concerns regarding the stability and generalizability of ratings at the item level. Future research involving mirrored versions or subtle perceptual differences should therefore consider recruiting a greater number of participants per image version to enhance the robustness and representativeness of individual stimulus ratings. The observed plausibility difference likely reflects subtle lighting asymmetries combined with observers' inherent preference for the left side of the face (Lindell, 2013 ; Powell & Schirillo, 2009 ). Future research might explore under what conditions such lateralization effects significantly influence plausibility without affecting recognition accuracy or perceived arousal. Conclusions This study presents a cost-effective pipeline for generating high-quality emotional stimuli from neutral facial images while preserving individual identity. The synthesized expressions—reward, affiliative, and dominance smiles, disgust, and neutral—were validated across emotional content, arousal, and plausibility. Notably, the confusions between affiliative and dominance smiles underscore the interpretive complexity of socially nuanced expressions. Together, these findings highlight the challenges of perceiving subtle affective signals and offer a validated, scalable tool to support future research on facial emotion recognition and social communication. Declarations Funding This research was supported by the China Scholarship Council under Grant No. 202208210077. Conflicts of interest/Competing interests The authors declare no competing financial or non-financial interests related to this work. Ethics approval The study protocol was approved by the Ethics Committee of the Department of Psychology, Humboldt-Universität zu Berlin (Approval No. 2024-17). All procedures complied with the Declaration of Helsinki. Consent to participate All participants provided informed consent electronically before beginning the online survey. Consent for publication Participants provided informed consent for the use and publication of their anonymized rating data. Availability of data and materials The study was preregistered on OSF. The preregistration and all anonymized rating data are available at https://osf.io/j4dfr/?view_only=8f6a7b25d26f4b79a6764eb398594285 (Open-Ended Registration, registered on August 7, 2024). The full set of in-house-derived stimuli is available upon request from the corresponding author. Any materials originating from external databases remain available via their original owners due to redistribution restrictions. Authors’ contributions Jin Gao: Conceptualization; data curation; formal analysis; investigation; methodology; software; visualization; writing – original draft; writing – review and editing. Werner Sommer (Prof.): Conceptualization; funding acquisition; methodology; project administration; resources; supervision; writing – review and editing. Rasha Abdel Rahman (Prof.): Conceptualization; funding acquisition; supervision; writing – review and editing. Wei-Jun Li (Prof.): Supervision; writing – review and editing. Public Significance Statements Not every smile signals happiness—some build rapport, others assert status. We offer a rigorous yet easy-to-use pipeline, built with specialized face-editing software, that turns neutral photos into controlled sets of reward, affiliative, and dominance smiles and validates them for recognition accuracy, arousal, and plausibility, giving researchers a clear recipe for studying nuanced emotion signals. The results reveal shared cues but frequent confusions among these smiles, reminding the public that even tiny facial shifts shape what we see and convey. References Bänziger, T., Mortillaro, M., & Scherer, K. R. (2012). Introducing the Geneva Multimodal Expression Corpus for experimental research on emotion perception. Emotion , 12 (5), 1161–1179. https://doi.org/10.1037/a0025827 Behrouzi, T., Shahroudnejad, A., & Mousavi, P. (2025). Maskrenderer: 3D-infused multi-mask realistic face reenactment. Pattern Recognition , 157 , 110891. https://doi.org/10.1016/j.patcog.2024.110891 Bijsterbosch, G., Mobach, L., Verpaalen, I. A. M., Bijlstra, G., Hudson, J. L., Rinck, M., & Klein, A. M. (2021). Validation of the child models of the Radboud Faces Database by children. International Journal of Behavioral Development , 45 (2), 146–152. https://doi.org/10.1177/0165025420935631 Bishara, A. J., & Hittner, J. B. (2012). Testing the significance of a correlation with nonnormal data: Comparison of Pearson, Spearman, transformation, and resampling approaches. Psychological Methods , 17 (3), 399–417. https://doi.org/10.1037/a0028087 Borod, J. C., Haywood, C. S., & Koff, E. (1997). Neuropsychological aspects of facial asymmetry during emotional expression: A review of the normal adult literature. Neuropsychology Review , 7 (1), 41–60. https://doi.org/10.1007/BF02876972 Busin, Y., Lukasova, K., Asthana, M. K., & Macedo, E. C. (2018). Hemiface differences in visual exploration patterns when judging the authenticity of facial expressions. Frontiers in Psychology , 8 , 2332. https://doi.org/10.3389/fpsyg.2017.02332 Crookes, K., Ewing, L., Gildenhuys, J., Kloth, N., Hayward, W. G., Oxner, M., Pond, S., & Rhodes, G. (2015). How well do computer-generated faces tap face expertise? PLOS ONE , 10 (11), e0141353. https://doi.org/10.1371/journal.pone.0141353 Dawel, A., Miller, E. J., Horsburgh, A., & Ford, P. (2022). A systematic survey of face stimuli used in psychological research 2000–2020. Behavior Research Methods , 54 (4), 1889–1901. https://doi.org/10.3758/s13428-021-01705-3 DeBruine, L., & Jones, B. (2017). Face Research Lab London Set [Dataset]. Figshare . https://doi.org/10.6084/m9.figshare.5047666.v1 Delicato, L. S. (2020). A robust method for measuring an individual’s sensitivity to facial expressions. Attention Perception & Psychophysics , 82 (6), 2924–2936. https://doi.org/10.3758/s13414-020-02043-w Diel, A., Weigelt, S., & Macdorman, K. F. (2021). A meta-analysis of the uncanny valley’s independent and dependent variables. Journal of Human-Robot Interaction , 11 (1). Article 1. https://doi.org/10.1145/3470742 Dyck, M., Winbeck, M., Leiberg, S., Chen, Y., Gur, R. C., & Mathiak, K. (2008). Recognition profile of emotions in natural and virtual faces. PLOS ONE , 3 (11), e3628. https://doi.org/10.1371/journal.pone.0003628 Ebner, N. C., Riediger, M., & Lindenberger, U. (2010). FACES—A database of facial expressions in young, middle-aged, and older women and men: Development and validation. Behavior Research Methods , 42 (1), 351–362. https://doi.org/10.3758/BRM.42.1.351 Ekman, P. (1976). Pictures of facial affect . Consulting Psychologists. Ekman, P., & Friesen, W. V. (1982). Felt, false, and miserable smiles. Journal of Nonverbal Behavior , 6 (4), 238–252. https://doi.org/10.1007/BF00987191 Ekman, P., & Rosenberg, E. L. (1997). What the Face Reveals: Basic and Applied Studies of Spontaneous Expression Using the Facial Action Coding System (FACS) . Oxford University Press. Fleiss, J. L. (1971). Measuring nominal scale agreement among many raters. Psychological Bulletin , 76 (5), 378–382. https://doi.org/10.1037/h0031619 Goeleven, E., De Raedt, R., Leyman, L., & Verschuere, B. (2008). The Karolinska Directed Emotional Faces: A validation study. Cognition and Emotion , 22 (6), 1094–1118. https://doi.org/10.1080/02699930701626582 Hareli, S., & Hess, U. (2012). The social signal value of emotions. Cognition and Emotion , 26 (3), 385–389. https://doi.org/10.1080/02699931.2012.665029 Hess, U., Senécal, S., Kirouac, G., Herrera, P., Philippot, P., & Kleck, R. E. (2000). Emotional expressivity in men and women: Stereotypes and self-perceptions. Cognition and Emotion , 14 (5), 609–642. https://doi.org/10.1080/02699930050117648 Horstmann, G. (2003). What do facial expressions convey: Feeling states, behavioral intentions, or action requests? Emotion , 3 (2), 150–166. https://doi.org/10.1037/1528-3542.3.2.150 Kammoun, A., Slama, R., Tabia, H., Ouni, T., & Abid, M. (2022). Generative adversarial networks for face generation: A survey. ACM Computing Surveys , 55 (5), 94:1–9437. https://doi.org/10.1145/3527850 Kaulard, K., Cunningham, D. W., Bülthoff, H. H., & Wallraven, C. (2012). The MPI facial expression database—A validated database of emotional and conversational facial expressions. PLOS ONE , 7 (3), e32321. https://doi.org/10.1371/journal.pone.0032321 Klingner, C. M., & Guntinas-Lichius, O. (2023). Facial expression and emotion. Laryngo-Rhino-Otologie , 102 (Suppl 1), S115–S125. https://doi.org/10.1055/a-2003-5687 Koo, T. K., & Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of Chiropractic Medicine , 15 (2), 155–163. https://doi.org/10.1016/j.jcm.2016.02.012 Kring, A. M., & Gordon, A. H. (1998). Sex differences in emotion: Expression, experience, and physiology. Journal of Personality and Social Psychology , 74 (3), 686–703. https://doi.org/10.1037/0022-3514.74.3.686 Krumhuber, E. G., Kappas, A., & Manstead, A. S. R. (2013). Effects of dynamic aspects of facial expressions: A review. Emotion Review , 5 (1), 41–46. https://doi.org/10.1177/1754073912451349 Krumhuber, E. G., Tamarit, L., Roesch, E. B., & Scherer, K. R. (2012). FACSGen 2.0 animation software: Generating three-dimensional FACS-valid facial expressions for emotion research. Emotion , 12 (2), 351–363. https://doi.org/10.1037/a0026632 Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics , 33 (1), 159–174. https://doi.org/10.2307/2529310 Lang, P. J. (1980). Behavioral treatment and bio-behavioral assessment: Computer applications. In J. B. Sidowski, J. H. Johnson, & T. A. Williams (Eds.), Technology in mental health care delivery systems (pp. 119–137). Ablex. Lindell, A. K. (2013). The silent social/emotional signals in left and right cheek poses: A literature review. Laterality , 18 (5), 612–624. https://doi.org/10.1080/1357650X.2012.737330 Ma, D. S., Correll, J., & Wittenbrink, B. (2015). The Chicago face database: A free stimulus set of faces and norming data. Behavior Research Methods , 47 (4), 1122–1135. https://doi.org/10.3758/s13428-014-0532-5 Martin, J., Rychlowska, M., Wood, A., & Niedenthal, P. M. (2017). Smiles as multipurpose social signals. Trends in Cognitive Sciences , 21 (11), 864–877. https://doi.org/10.1016/j.tics.2017.08.007 Mehu, M., Mortillaro, M., Bänziger, T., & Scherer, K. R. (2012). Reliable facial muscle activation enhances recognizability and credibility of emotional expression. Emotion , 12 (4), 701–715. https://doi.org/10.1037/a0026632 Mollahosseini, A., Hasani, B., & Mahoor, M. H. (2019). AffectNet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing , 10 (1), 18–31. https://doi.org/10.1109/TAFFC.2017.2740923 Niedenthal, P. M., Mermillod, M., Maringer, M., & Hess, U. (2010). The simulation of smiles (SIMS) model: Embodied simulation and the meaning of facial expression. Behavioral and Brain Sciences , 33 (6), 417–433. https://doi.org/10.1017/S0140525X10000865 Olszanowski, M., Pochwatko, G., Kuklinski, K., Scibor-Rylski, M., Lewinski, P., & Ohme, R. K. (2015). Warsaw set of emotional facial expression pictures: A validation study of facial display photographs. Frontiers in Psychology , 5 , 1516. https://doi.org/10.3389/fpsyg.2014.01516 Plant, E. A., Hyde, J. S., Keltner, D., & Devine, P. G. (2000). The gender stereotyping of emotions. Psychology of Women Quarterly , 24 (1), 81–92. https://doi.org/10.1111/j.1471-6402.2000.tb01024.x Pochedly, J. T., Widen, S. C., & Russell, J. A. (2012). What emotion does the facial expression of disgust express? Emotion , 12 (6), 1315–1319. https://doi.org/10.1037/a0027998 Powell, W. R., & Schirillo, J. A. (2009). Asymmetrical facial expressions in portraits and hemispheric laterality: A literature review. Laterality , 14 (6), 545–572. https://doi.org/10.1080/13576500802680336 Pumarola, A., Agudo, A., Martinez, A. M., Sanfeliu, A., & Moreno-Noguer, F. (2020). GANimation: One-shot anatomically consistent facial animation. International Journal of Computer Vision , 128 (3), 698–713. https://doi.org/10.1007/s11263-019-01210-3 Rai, A., Gupta, H., Pandey, A., Carrasco, F. V., Takagi, S. J., Aubel, A., Kim, D., Prakash, A., & De La Torre, F. (2024). Towards realistic generative 3D face models. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (pp. 3726–3736). IEEE. https://doi.org/10.1109/WACV57701.2024.00370 Rellecke, J., Sommer, W., & Schacht, A. (2012). Does processing of emotional facial expressions depend on intention? Time-resolved evidence from event-related brain potentials. Biological Psychology , 90 (1), 23–32. https://doi.org/10.1016/j.biopsycho.2012.02.002 Roesch, E., Tamarit, L., Reveret, L., Grandjean, D., Sander, D., & Scherer, K. R. (2011). FACSGen: A tool to synthesize emotional facial expressions through systematic manipulation of facial action units. Journal of Nonverbal Behavior , 35 (1), 1–16. https://doi.org/10.1007/s10919-010-0095-9 Rozin, P., Lowery, L., & Ebert, R. (1994). Varieties of disgust faces and the structure of disgust. Journal of Personality and Social Psychology , 66 (5), 870–881. https://doi.org/10.1037/0022-3514.66.5.870 R Core Team. (2024). R: A language and environment for statistical computing . R Foundation for Statistical Computing. Rychlowska, M., Jack, R. E., Garrod, O. G. B., Schyns, P. G., Martin, J. D., & Niedenthal, P. M. (2017). Functional smiles: Tools for love, sympathy, and war. Psychological Science , 28 (9), 1259–1270. https://doi.org/10.1177/0956797617706082 Sauter, D. A. (2010). More than happy: The need for disentangling positive emotions. Current Directions in Psychological Science , 19 (1), 36–40. https://doi.org/10.1177/0963721409359290 Schmidtmann, G., Jennings, B. J., Sandra, D. A., Pollock, J., & Gold, I. (2020). The McGill Face Database: Validation and insights into the recognition of facial expressions of complex mental states. Perception , 49 (3), 310–329. https://doi.org/10.1177/0301006620901671 Shrout, P. E., & Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin , 86 (2), 420–428. https://doi.org/10.1037/0033-2909.86.2.420 Strohminger, N., Gray, K., Chituc, V., Heffner, J., Schein, C., & Heagins, T. B. (2016). The MR2: A multi-racial, mega-resolution database of facial stimuli. Behavior Research Methods , 48 (3), 1197–1204. https://doi.org/10.3758/s13428-015-0641-9 Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., & Ortega-Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. Information Fusion , 64 , 131–148. https://doi.org/10.1016/j.inffus.2020.06.014 Tottenham, N., Tanaka, J. W., Leon, A. C., McCarry, T., Nurse, M., Hare, T. A., Marcus, D. J., Westerlund, A., Casey, B., & Nelson, C. (2009). The NimStim set of facial expressions: Judgments from untrained research participants. Psychiatry Research , 168 (3), 242–249. https://doi.org/10.1016/j.psychres.2008.05.006 Tracy, J. L., Robins, R. W., & Schriber, R. A. (2009). Development of a FACS-verified set of basic and self-conscious emotion expressions. Emotion , 9 (4), 554–559. https://doi.org/10.1037/a0015766 Wagner, H. L. (1993). On measuring performance in category judgment studies of nonverbal behavior. Journal of Nonverbal Behavior , 17 (1), 3–28. https://doi.org/10.1007/BF00987006 Zhao, Z., Yaoma, K., Wu, Y., Burns, E., Sun, M., & Ying, H. (2024). Other ethnicity effects in ensemble coding of facial expressions. Attention Perception & Psychophysics . https://doi.org/10.3758/s13414-024-02920-8 Footnotes Please see https://osf.io/xh298 for more details. Please see https://github.com/JinGao1997/Facial-Expressions-Rating-Task for more details Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 19 Mar, 2026 Read the published version in Psychological Research → Version 1 posted Editorial decision: Revision requested 21 Oct, 2025 Reviews received at journal 15 Oct, 2025 Reviewers agreed at journal 24 Jul, 2025 Reviews received at journal 23 Jul, 2025 Reviewers agreed at journal 15 Jul, 2025 Reviewers invited by journal 23 Jun, 2025 Editor assigned by journal 18 Jun, 2025 Submission checks completed at journal 17 Jun, 2025 First submitted to journal 16 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6908842","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":475340325,"identity":"cc965d15-319e-4ec7-9b1a-e9a5b1feb76c","order_by":0,"name":"Jin Gao","email":"","orcid":"","institution":"Humboldt-Universität zu Berlin","correspondingAuthor":false,"prefix":"","firstName":"Jin","middleName":"","lastName":"Gao","suffix":""},{"id":475340327,"identity":"9ca49380-ec05-4737-9493-b19b57b9040d","order_by":1,"name":"Werner Sommer","email":"data:image/png;base64,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","orcid":"","institution":"Humboldt-Universität zu Berlin","correspondingAuthor":true,"prefix":"","firstName":"Werner","middleName":"","lastName":"Sommer","suffix":""},{"id":475340329,"identity":"856c57bc-a5b0-46fe-9f29-2ab9a79f4737","order_by":2,"name":"Rasha Abdel Rahman","email":"","orcid":"","institution":"Humboldt-Universität zu Berlin","correspondingAuthor":false,"prefix":"","firstName":"Rasha","middleName":"Abdel","lastName":"Rahman","suffix":""},{"id":475340333,"identity":"27b1433d-81d9-410e-91d2-cc78a1cec345","order_by":3,"name":"Wei-Jun Li","email":"","orcid":"","institution":"Liaoning Normal University","correspondingAuthor":false,"prefix":"","firstName":"Wei-Jun","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2025-06-16 23:08:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6908842/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6908842/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00426-026-02263-z","type":"published","date":"2026-03-19T15:58:29+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":85864034,"identity":"fd7f7fb8-96f6-4a82-8767-5b95ab6fe125","added_by":"auto","created_at":"2025-07-02 12:46:51","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":282282,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePipeline for Generating Facial Expression Stimuli\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e The final output includes five expressions: neutral (centre), reward smile (top left), affiliative smile (bottom left), dominance smile (top right), and disgust (bottom right).\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6908842/v1/7413f3991266982c5a1cf891.png"},{"id":85865707,"identity":"013fc7e8-f279-47d9-957e-cdd8bd5bb785","added_by":"auto","created_at":"2025-07-02 13:02:52","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":233274,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eExample of the Evaluation Process for a Single Image\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e Participants rated each image on three dimensions presented sequentially: (1) emotional content, (2) arousal, and (3) plausibility. Each question was displayed below the image.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6908842/v1/7676cbcd17dff2a542fe316d.png"},{"id":85864038,"identity":"e9780311-7cbe-4a9a-a673-d3322da25eaa","added_by":"auto","created_at":"2025-07-02 12:46:52","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":261839,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eConfusion Matrix: Percentage of Chosen Emotions Per Intended Emotional Expression\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e The diagonal elements represent correct classifications, while the off-diagonal elements show the distribution of misclassifications. The colour intensity corresponds to the percentage values, with brighter colours indicating higher percentages. Chance level is 16.7%, marked by a red arrow.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6908842/v1/fc02ba74237747d907204e01.png"},{"id":85865281,"identity":"6a5b6a47-52b2-432a-968f-1970332ee7ae","added_by":"auto","created_at":"2025-07-02 12:54:52","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":497590,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eExpressor Ranking for Combined UHR and Plausibility Scores with Group-Wise Gender Distribution\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e Combined scores were derived by averaging the standardized arcsine-transformed unbiased hit rates (UHR) (across all expressions) and plausibility ratings. The figure ranks all 90 expressors, divided into ten subgroups of nine expressors each, based on their combined scores. Shaded background sections on the bar chart indicate the rank ranges for each subgroup. The horizontal bar chart shows the combined scores (x-axis) and model names (y-axis), while the pie charts illustrate the gender distribution within each group, with blue representing male expressors and pink representing female expressors.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6908842/v1/7ba96a4ea89d1417f100200c.png"},{"id":105223328,"identity":"73a7b9e2-e6ba-4d0e-a0a8-925847126a2f","added_by":"auto","created_at":"2026-03-23 16:04:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2267626,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6908842/v1/2d701f37-7c1f-400d-9d3d-923f516924e1.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Shades of Smiles: Creating Variants of Smiles from Neutral Images of Real Individuals - Method and Validation","fulltext":[{"header":"Introduction","content":"\u003cp\u003eFacial expressions are a valuable research tool for the study of emotions, as they convey a large range of affects states, including feelings, intentions, and desires (Horstmann, 2003). They can be precisely described by observable facial surface changes driven by specific patterns of muscle activity, referred to as Action Units (Ekman \u0026amp; Rosenberg, 1997; Klingner \u0026amp; Guntinas-Lichius, 2023). Because of these attributes, researchers have collected various facial expression databases providing standardized stimuli. However, acquiring images of non-standard emotional expressions across a sufficiently diverse pool of expressors remains challenging. In this study, we suggest a solution based on the artificial creation of custom-defined expressions from photographs of individuals displaying neutral expressions.\u003c/p\u003e\n\u003cp\u003eUnless cartoon faces are used, most realistic facial emotion stimuli have been sourced from real human expressors (Ekman, 1976; Olszanowski et al., 2015; Tracy et al., 2009). For a review, see Dawel et al., (2022). However, integrating large datasets with high-quality expression stimuli presents significant challenges. For instance, AffectNet\u0026mdash;a database of facial affect compiled from the internet using numerous emotion-related search queries\u0026mdash;comprises over one million images with faces and facial landmark points (Mollahosseini et al., 2019). Nonetheless, these images are typically captured under uncontrolled conditions with considerable variability in lighting, pose, resolution, and background, and only a fraction of images depict a given emotion accurately. Specifically, only about 48% of images queried for \u0026quot;happy\u0026quot; expressions were annotated as happiness; percentages for other emotions were even lower (below 20%; see Mollahosseini et al., 2019, Table 4). In contrast, databases like the MPI Facial Expression Database (Kaulard et al., 2012) and the McGill Face Database (Schmidtmann et al., 2020) provide highly controlled stimuli obtained from trained models. However, training these expressors to reliably produce specific Action Units (AUs) or nuanced expressions\u0026mdash;such as shame or composite expressions like a sad smile\u0026mdash;requires considerable time and effort, and not all posers achieve consistent results (Mehu et al., 2012; Tottenham et al., 2009). For instance, the MPI Facial Expression Database includes an impressive array of 55 natural emotional and conversational expressions elicited through a method-acting protocol, in which participants were provided with specific everyday scenarios designed to evoke natural and authentic expressions; however, it comprises only 19 German models (Kaulard et al., 2012) Similarly, the McGill Face Database features 93 expressions of mental states but is based on merely two professional expressors (Schmidtmann et al., 2020). Consequently, for specialized research or specific applications, even these extensive databases are limited.\u003c/p\u003e\n\u003cp\u003eAdvances in generative adversarial networks (GANs) and related AI-driven methods have substantially improved the generation of high-quality facial stimuli (Kammoun et al., 2022). For instance, Deepfake technologies, based on GAN architectures, can produce highly realistic face images that are often indistinguishable from real photographs (Tolosana et al., 2020). Similarly, GANimation (Pumarola et al., 2020) offers the ability to manipulate facial Action Units (AUs) to generate expressions in a controllable manner. However, while both methods excel when the input closely matches the training data distribution, they tend to falter when applied to faces with distinctive structures or when generating subtle, nuanced emotional expressions. In particular, Deepfake technologies, although highly convincing in overall face appearance, often lack precise control over specific emotional expressions. Likewise, outputs from GANimation can appear distorted or unrecognizable when dealing with complex or unfamiliar facial features. These limitations, rooted in training biases, restrict the applicability of these methods in research contexts that require a large set of subtle, systematically varied emotional expressions across different individuals.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFacial modelling and animation techniques from computer vision and graphics provide another solution for generating extensive sets of emotional facial expressions. Unlike physics-based simulations that rely on complex geometric and muscular representations (Behrouzi et al., 2025; Rai et al., 2024), the FaceGen software suite\u0026mdash;particularly FACSGen\u0026mdash;offers an accessible and intuitive platform for AU manipulation. FACSGen 2.0 integrates advanced algorithms with a user-friendly interface to map photorealistic textures onto virtual faces, enabling the precise generation of both static and dynamic expressions. Krumhuber et al., (2012) demonstrated the utility of this software in producing AU-defined facial expressions, including canonical emotions (e.g., happiness, anger, fear) and less common social emotions (e.g., embarrassment, contempt), across various genders and ethnicities using grayscale faces.\u003c/p\u003e\n\u003cp\u003eBuilding on the work of Krumhuber et al. (2012), the present study focuses on three specific types of smiles\u0026mdash;reward, affiliative, and dominance\u0026mdash;that commonly function as social signals in interpersonal contexts (Hareli \u0026amp; Hess, 2012). These smiles are defined by distinct Action Unit (AU) configurations \u0026nbsp;(Martin et al., 2017; Niedenthal et al., 2010). Reward smiles are displayed to express happiness, reward others or to communicate positive experiences or intentions; they involve the activation of the Lip Corner Puller (AU12), Inner-Outer Brow Raiser (AU1-2), Sharp Lip Puller (AU13), and Dimpler (AU14) (Rychlowska et al., 2017). Affiliative smiles are shown to signal appeasement and create and maintain social bonds and also involve AU12 but are distinguished from Reward smiles by activating the Lip Pressor (AU24) and Dimpler (AU14). Dominance smiles are displayed to negotiate status within and across social hierarchies; they also feature AU12 but asymmetrically, involving either the right or left Lip Corner Puller (AU12R or AU12L). They are also characterized by the Upper Lid Raiser (AU5), Nose Wrinkler (AU9), Cheek Raiser (AU6), and Upper Lip Raiser (AU10). To allow for manipulation checks and to differentiate positive from at least one negative expression, we also included disgust expressions. Disgust, a negative expression, shares the Nose Wrinkler (AU9) with the dominance smile (Pochedly et al., 2012). Furthermore, neutral expressions were derived from the model faces with the same pipeline but without AU activation, allowing these expressions to serve as experimental baseline.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile Krumhuber et al., (2012) pioneered AU-based facial expression generation using FACSGen 2.0 , the present study addresses key limitations in their study and extends it to broader applications. First, whereas Krumhuber et al., (2012) used a small number of expressors (at most four expressors per experimental condition), our stimuli were derived from a much broader range of real front-facing neutral portraits sourced from multiple databases. These images were converted into standardized head expressors while preserving identity-specific features. Although this study focused on a single racial group to ensure experimental control, our method is scalable to more diverse populations. Second, although Krumhuber et al. (2012) included a few less-common social expressions (e.g., embarrassment, contempt), their work did not provide a comprehensive pipeline for generating and validating nuanced, socially meaningful variants of expressions. In contrast, our integrated pipeline systematically produces and validates three subtle smile types\u0026mdash;reward, affiliative, and dominance\u0026mdash;each capturing distinct social signals such as appeasement versus assertiveness. Third, building on the validation approach used by Krumhuber et al., (2012), we evaluated recognition accuracy, perceived arousal, and plausibility. Crucially, we introduced the arcsine-transformed Unbiased Hit Rate (UHR), a response-bias-resistant metric, and incorporated it into a composite index combining plausibility ratings, allowing for a more comprehensive assessment of model performance. Finally, to support reproducibility, we provide open-source validation code and share all stimuli derived from our in-house images.\u003c/p\u003e\n\u003cp\u003eWe hypothesised that 1) participants would reliably recognize and accurately identify the distinct emotional expressions depicted in the synthesized facial images, that is, \u0026quot;reward smiles,\u0026quot; \u0026quot;affiliative smiles,\u0026quot; \u0026quot;dominance smiles,\u0026quot; \u0026quot;disgust,\u0026quot; and \u0026quot;neutral expressions.\u0026quot; The accuracy rates for identifying these expressions were expected to be significantly better than chance, indicating clear and distinguishable emotional contents. 2) Smiles and disgust were expected to be associated with higher perceived arousal levels than neutral expressions. 3) The synthesized images were expected to be perceived as highly plausible, with participants judging the images as being derived from photographs of real humans.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eConstruction Stage\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo create the emotional expression stimuli, we followed a structured series of steps as illustrated in Fig. 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStep 1: selection and preparation of database sources.\u0026nbsp;\u003c/strong\u003eAs original images we selected 90 high-resolution front-facing photographs of Caucasian individuals with neutral facial expressions from several sources (DeBruine \u0026amp; Jones, 2017; Ebner et al., 2010; Ma et al., 2015; Rellecke et al., 2012; Strohminger et al., 2015). Stimuli comprised 45 male and 45 female expressors, with mean ages of 25.7 and 26.3 years, respectively. While there were slight variations in parameters across picture sources (e.g., the colour of backgrounds and the size of the pictures), they were sufficiently similar to provide a consistent basis for the generation of head models since they were, taken under similar shooting conditions and ensured clear visibility of individual facial features. Besides, we selected only models without beards or non-face features, such as, scars, tattoos or piercings, jewellery, glasses, makeup or any clothing that covered the neck. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStep 2: generation of head models.\u0026nbsp;\u003c/strong\u003eWe utilized the FaceGen Modeller, a part of the FaceGen software suite (Roesch et al., 2011), allowing to generate a virtually infinite number of 3D facial identities as head model. The selected photographs were imported into FaceGen Modeller software. The software extracted facial geometry and skin texture information from these images, constructing detailed 3D head models with realistic skin textures and individual-specific features. Specifically, photorealistic skin texture was mapped onto the face. Different texture layers (i.e., diffuse colour, ambient occlusion, and gloss and normal maps) were also included, which are combined during the rendering stage to achieve the final appearance (Krumhuber et al., 2012).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStep 3: generating and adjusting facial expressions based on action units (AUs).\u0026nbsp;\u003c/strong\u003eFACSGen 2.0 software (Krumhuber et al., 2012) allows detailed adjustment and combining various facial expressions based on Action Units (AUs), enabling researchers to simulate a range of emotional expressions without altering the identity of the expressors. We began the process by selecting specific combinations of AUs that correspond to the targeted facial expressions \u0026quot;reward smile,\u0026quot; \u0026quot;affiliative smile,\u0026quot; \u0026quot;dominance smile,\u0026quot; and \u0026quot;disgust.\u0026quot; The AU combinations for each expression (see Table 1) were selected based on the smiling prototypes described by Martin et al. (2017) and the disgust prototype (Pochedly et al. 2012). Additionally, since all original portraits displayed neutral expressions, the \u0026apos;neutral expression\u0026apos; of the 3D models was directly generated from the imported photographs without any manipulation of AUs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"565\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eExpression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003eAU combinations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003ereward smile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAU1 + AU2 +\u003c/strong\u003e AU6 + AU7 +\u003cstrong\u003e\u0026nbsp;AU12 + AU13 + AU14 +\u003c/strong\u003e AU20 + AU26 + ahh + i + k\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eaffiliative smile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003eAU6 + AU7 +\u003cstrong\u003e\u0026nbsp;AU12 +\u003c/strong\u003e AU13 +\u003cstrong\u003e\u0026nbsp;AU14 +\u003c/strong\u003e AU20 +\u003cstrong\u003e\u0026nbsp;AU24\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003edominance smile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003eAU2 + AU4 + (AU1+2+4) +\u003cstrong\u003e\u0026nbsp;AU5 + AU6 +\u003c/strong\u003e AU7 +\u003cstrong\u003e\u0026nbsp;AU9 + AU10 + AU12+\u003c/strong\u003e AU14_Unilateral + AU25 + AU17\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003edisgust\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 450px;\"\u003e\n \u003cp\u003eAU4 + (AU1+4) +\u003cstrong\u003e\u0026nbsp;AU6 +\u003c/strong\u003e AU7 + AU9+\u003cstrong\u003e\u0026nbsp;AU10 +\u003c/strong\u003e AU11+ AU16 + AU15 + AU17 + AU20 + AU24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eTargeted Facial Expressions for Each Emotion and Their Corresponding Action Units (AUs)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNotes.\u003c/em\u003e AUs in boldface are considered essential for inclusion in the final portrayal of a given expression; AUs in normal typeface can occur at any level of intensity. AUs named \u0026ldquo;ahh\u0026rdquo;, \u0026ldquo;I\u0026rdquo;, \u0026ldquo;k\u0026rdquo;, and those in parentheses are pre-defined combinations of AUs in FACSGen 2.0, which can be used to fine-tune the expressions to better approximate their prototypes.\u003c/p\u003e\n\u003cp\u003eIn FACSGen 2.0, each AU is represented by a slider controlling the magnitude of the morph target, which gradually transforms the original neutral expression (0%) into the fully activated target expression (100%). We used the slider-based controls to adjust the intensity of the AUs specified for the target emotion, fine-tuning them to closely match the prototype of each specific expression. This was necessary because setting the AU activity at a preset level for all faces may induce artificial looking results. In adjusting the AU strengths we followed the principle of keeping the expressions as natural as possible without unnatural facial distortions, such as teeth-lip intersections or excessive muscle activation. For AU12\u0026mdash;activated in all three types of smiles and playing a crucial role in defining the characteristic look of each smile\u0026mdash;the intensity was uniformly set to \u0026quot;80%\u0026quot; for all three smiles (without manual fine-tuning). It is important to note that FACSGen does not provide asymmetric AU12 activation, which is considered to be crucial for the dominance smile (Martin et al., 2017; Niedenthal et al., 2010). To achieve this effect, we employed a compensatory method by combining \u0026quot;dragging one side of the mouth corner upward in FaceGen Modeller on the face picture\u0026quot; and increasing the intensity of AU14_Unilateral in FACSGen to approximate the asymmetric AU12 effect.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStep 4: standardization of display parameters and generation of static images.\u0026nbsp;\u003c/strong\u003eThe original hairstyle and hair colour of the individuals in the photographs cannot be retained during the 3D model generation process because the FACSGen 2.0 software does not support importing or replicating external features such as hairstyle and hair colour. To further standardize the images, we ensured uniformity of hairstyle and hair colour of the generated head models. Note, there are many hair styles and colours available in FACSGen that could be used. \u0026nbsp;For present purposes, all individuals of a given sex were assigned the same sex-typical hairstyle and all models, regardless of sex, were given a uniformly brunette hair colour. Additionally, the eye colours of the models were set according to the frequencies typically observed in Caucasian populations, that is, 40% brown, 30% blue, 20% green and 8% grey (Sulem et al., 2007). Display parameters, including frontal face orientation, taupe background, and lighting conditions, were uniformly set for all 3D models. To simplify evaluation and communication, we exported 2D front view projections of the models at a resolution of 800 \u0026times; 1200 pixels with 32-bit depth. The 2D projections captured the key features of the 3D models in a single, clear image, making it straightforward to identify essential characteristics and potential issues. These images were then used for validation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eValidation Stage\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants.\u003c/strong\u003e As detailed in our pre‐registration\u003ca href=\"#_ftn1\" name=\"_ftnref1\" title=\"\"\u003e\u003c/a\u003e\u003csup\u003e1\u003c/sup\u003e，we conducted an a priori power analysis in G*Power 3.1 for our mixed‐design ANOVA (within‐between interaction) with the following parameters: medium effect size (\u003cem\u003ef\u003c/em\u003e = 0.25), \u003cem\u003e\u0026alpha;\u003c/em\u003e = .01, power (1 \u0026minus; \u003cem\u003e\u0026beta;\u003c/em\u003e) = .90, total groups = 10 (5 lists \u0026times; 2 versions), repeated measures = 5 (emotion categories), assumed correlation among repeated measures = .50, and non-sphericity correction (\u003cem\u003e\u0026epsilon;\u003c/em\u003e) = 1. This yielded a minimum of 80 participants (8 per group). To enhance the precision of item‐level variance and confidence intervals, we increased to \u003cem\u003eN\u003c/em\u003e = 130 raters (13 per group), exceeding the a priori requirement and ensuring robust power for all planned contrasts.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAccordingly, via Prolific (https://www.prolific.com/), we initially recruited \u003cem\u003eN\u003c/em\u003e = 135 participants (paid \u0026pound;7.62/hr). All were right‐handed adults (18\u0026ndash;34 years) with normal or corrected vision, no reported physical or mental health issues, and self‐identified as White to preclude other‐race effects given our Caucasian models (Crookes et al., 2015; Zhao et al., 2024). The study adhered to the Declaration of Helsinki and received ethics approval from the Department of Psychology at Humboldt-Universit\u0026auml;t zu Berlin. Participants gave informed consent, including agreement to non-dissemination of images. After excluding one rater who reported \u0026ldquo;Other\u0026rdquo; gender and four with below‐chance accuracy, the final sample comprised 130 participants: 70 women (\u003cem\u003eM\u0026nbsp;\u003c/em\u003e= 25.73 years, \u003cem\u003eSD\u003c/em\u003e = 4.30) and 60 men (\u003cem\u003eM\u003c/em\u003e = 25.68 years, \u003cem\u003eSD\u003c/em\u003e = 5.02).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExperimental stimuli.\u0026nbsp;\u003c/strong\u003eThe images used had been synthesized from portraits of 90 expressors (45 males and 45 females) with neutral expression (according to the normative data of the sources). For each portrait, five images with different expressions were created: \u0026quot;reward smile,\u0026quot; \u0026quot;affiliative smile,\u0026quot; \u0026quot;dominance smile,\u0026quot; \u0026quot;disgust,\u0026quot; and \u0026quot;neutral expression.\u0026quot; For clarity in describing the assignment scheme, the synthesized emotional expressions are referred to as if they had been displayed by the expressors themselves, although they were generated from their neutral portraits. To control for potential hemiface effects\u0026mdash;a phenomenon referring to asymmetry biases in human emotion recognition, where there is preferential looking at the left hemiface when actively exploring face images (Borod et al., 1997; Busin et al., 2018), we created two mirrored versions of each facial image, labelled as L-version and R-version, respectively.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo minimize rater fatigue but ensure comprehensive coverage of all images, we divided the 90 model identities into five model groups of 18 (9 males and 9 females) each. A Latin Square design assigned each model in a group to display only one of five expressions (reward, affiliative, dominance, disgust, neutral) in the original or mirrored version, resulting in 5 different lists of images. For each of these five list a second list was created with mirrored versions of the images, yielding a total of 10 image lists. For example, in List 1, expressors in Groups 1\u0026ndash;5 displayed reward, neutral, disgust, dominance, and affiliative, respectively; in List 2, the same expressors displayed neutral, disgust, dominance, affiliative, and reward, respectively. This rotation continued through all ten lists; hence \u0026nbsp;within a list each model had presented only one expression but across lists each model had shown all expressions in original and mirrored versions. Since a given rater received only one of the 10 lists, we avoided crosstalk across different expressions in a given model within the same rater. Each of the ten lists was presented to a distinct group of 13 participants, such that every image was rated by 13 different individuals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eProcedure.\u0026nbsp;\u003c/strong\u003eThe ratings were conducted online using the SoSci Survey platform. Raters were informed that they were taking part in a validation study of emotional pictures of human faces. Before the main task, participants were instructed to adjust their viewing distance to their screens at home to standardize the visual angle of the facial images; the horizontal width of the face image should roughly correspond to the combined width across index, middle, and ring fingers with arms outstretched. This arrangement yields a visual angle of approximately 4\u0026deg; horizontally, as used in many experiments.\u003c/p\u003e\n\u003cp\u003eThe instructions emphasized that there are no \u0026apos;right\u0026apos; or \u0026apos;wrong\u0026apos; answers and participants were encouraged to base their responses on their spontaneous and subjective opinions but to pay close attention to the (sometimes) subtle differences between the facial expressions. During the practice phase, participants were familiarized with the six expression categories used in the task: reward smile, affiliative smile, dominance smile, neutral, disgust, and other. These categories were operationally defined to support consistent judgment, based on brief descriptions and everyday scenarios (e.g., reward smile: \u0026quot;expressing positive internal happiness, for example, after receiving good news\u0026quot;; affiliative smile: \u0026quot;inviting and maintaining mutually positive and beneficial social bonds, typically expressed when greeting a stranger\u0026quot;; dominance smile: \u0026quot;instantiating, stabilizing, or maintaining dominance in a social relation, for example, when disapproving the ideas or actions of a person viewed as intellectually or physically inferior\u0026quot;; disgust: \u0026quot;rejection or revulsion of something or someone potentially contagious, offensive, distasteful, or unpleasant, for example, seeing someone vomiting\u0026quot;; neutral expression: \u0026quot;a non-expressive posture of the face\u0026quot;; other expression: \u0026quot;not belonging to one of the five previous categories\u0026quot;).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs illustrated in Fig. 2, for each image, three questions were posed in order: 1) \u0026quot;Which facial expression does this person primarily show?\u0026quot; (emotional content): Here, the participants should choose one of six options: reward smile, affiliative smile, dominance smile, neutral, disgust, and other. The ordering of the expression labels was kept constant within each rater but was counterbalanced across participants. 2) \u0026quot;How strong is this person\u0026apos;s level of arousal?\u0026quot;. For these arousal ratings we used the Self-Assessment Manikin (SAM; P. Lang, 1980), a pictorial technique with graphic figures representing different levels of arousal, ranging from a completely relaxed, unstimulated, and calm state (depicted by a manikin with eyes closed) to a highly awake, stimulated, and excited state. 3) \u0026quot;How likely is it that this face is based on a real human photograph?\u0026quot; (plausibility). Raters selected one of seven probability levels, ranging from \u0026quot;Impossible\u0026quot; to \u0026quot;Certain\u0026quot;. After each answer, participants clicked the \u0026quot;Next\u0026quot; button to proceed to the subsequent question. For Questions 2 and 3, participants were encouraged to utilize the full range of the provided rating scales.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAfter completing several practice tasks, participants proceeded to the main experiment. Each image was presented along with the rating questions on the same screen, allowing participants to respond immediately after viewing the image. There was no time limit for completing the ratings, and participants were encouraged to take breaks whenever needed at their own discretion.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThroughout the experiment, five attention check questions were randomly interspersed, as suggested by Prolific, to ensure participants remained engaged and responded attentively. These questions explicitly requested to complete a task in a specific way, for example \u0026ldquo;The colour test you are about to take part is very simple, when asked for your favourite colour you must select \u0026apos;Red\u0026apos;\u0026rdquo;.\u003c/p\u003e"},{"header":"Data Analyses and Results","content":"\u003cp\u003eIn the following we report results for the emotional category ratings as simple hit rates (SHR), unbiased hit rates (UHR), and as confusion matrix. Then we report the results for the arousal and plausibility ratings, followed by an analysis of interrater reliability. All analyses were conducted using R Statistical Software (v4.3.3; R Core Team 2024).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSimple Hit Rate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSHR is a classic accuracy index, which is simple and intuitive (B\u0026auml;nziger et al., 2012). In the current study, SHR were mean percentages of correct responses indicating how often the chosen emotion agrees with the experimenter-intended emotional expression. SHR was calculated separately for each emotion category across all participants and expressors.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDescriptive statistics for SHR across expression categories are summarized in Table 2. The SHR was highest for neutral (\u003cem\u003eM =\u003c/em\u003e 92.6%, \u003cem\u003eSD =\u003c/em\u003e 11.3), reward (\u003cem\u003eM =\u003c/em\u003e 92.1%, \u003cem\u003eSD =\u003c/em\u003e 14.7), and disgust expressions (\u003cem\u003eM =\u003c/em\u003e 91.5%, \u003cem\u003eSD =\u003c/em\u003e 13.5), and substantially lower for affiliative (\u003cem\u003eM =\u003c/em\u003e 69.2%, \u003cem\u003eSD =\u003c/em\u003e 29.5) and dominance smiles (\u003cem\u003eM =\u003c/em\u003e 66.0%, \u003cem\u003eSD =\u003c/em\u003e 28.8). Given the six-alternative forced-choice format, the theoretical chance level was 16.7%. All mean SHR values were markedly above chance, suggesting that participants were able to recognize the expressions at rates well beyond guessing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"3\" cellpadding=\"0\" width=\"605\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eExpression Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003eSHR \u003cem\u003eM\u003c/em\u003e (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSHR \u003cem\u003eSD\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eUHR \u003cem\u003eM\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eUHR \u003cem\u003eSD\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eChance UHR \u003cem\u003eM\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003eChance UHR \u003cem\u003eSD\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eNeutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e92.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e11.300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.880\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.160\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e0.044\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eaffiliative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e69.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e29.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.594\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.305\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e0.038\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e0.016\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eDisgust\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e91.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e13.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.892\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.166\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003edominance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e66.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e28.800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.309\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e0.031\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e0.014\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eReward\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e92.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e14.700\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.798\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e0.219\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e0.049\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e0.016\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eMean (M) and Standard Deviation (SD) of Recognition Accuracy Measures by Expression Type\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003eNote\u003c/em\u003e. SHR = Simple Hit Rate (percentage of correct responses); UHR = Unbiased Hit Rate; Chance UHR is computed based on the marginal frequencies of the confusion matrix for each expression type (Wagner, 1993).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSHR data were analysed using a binomial (logit) generalised linear mixed model (GLMM). This model was chosen because SHR values are based on trial-level binary outcomes (correct/incorrect), which naturally follow a binomial distribution. Attempts to fit a linear mixed-effects model (LMM)\u0026mdash;even after arcsine transformation\u0026mdash;violated assumptions of normality and homoscedasticity. A GLMM thus provided a more appropriate framework for analysing the binary nature of the data while accounting for random variability at the rater level. In this model, expression category, group, rater gender, model gender, and version were specified as fixed effects, and rater (case) was included as a random intercept.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe model showed good overall fit (AIC = 8854.1, BIC = 8986.7). Relative to neutral expressions, both affiliative (\u003cem\u003e\u0026beta; =\u0026nbsp;\u003c/em\u003e\u0026ndash;1.95, \u003cem\u003eSE =\u0026nbsp;\u003c/em\u003e0.10, \u003cem\u003ez =\u0026nbsp;\u003c/em\u003e\u0026ndash;20.27, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001) and dominance smiles (\u003cem\u003e\u0026beta; =\u0026nbsp;\u003c/em\u003e\u0026ndash;2.12, \u003cem\u003eSE =\u003c/em\u003e 0.10, \u003cem\u003ez =\u0026nbsp;\u003c/em\u003e\u0026ndash;22.16, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001) were associated with significantly lower probabilities of a correct response. Disgust (\u003cem\u003e\u0026beta; =\u0026nbsp;\u003c/em\u003e\u0026ndash;0.16, \u003cem\u003eSE =\u0026nbsp;\u003c/em\u003e0.11, \u003cem\u003ez =\u0026nbsp;\u003c/em\u003e\u0026ndash;1.41, \u003cem\u003ep =\u0026nbsp;\u003c/em\u003e.16) and reward smiles (\u003cem\u003e\u0026beta; =\u0026nbsp;\u003c/em\u003e\u0026ndash;0.07, \u003cem\u003eSE =\u0026nbsp;\u003c/em\u003e0.11, \u003cem\u003ez =\u0026nbsp;\u003c/em\u003e\u0026ndash;0.58, \u003cem\u003ep =\u0026nbsp;\u003c/em\u003e.56) did not differ significantly from neutral. No significant effects were found for group, rater gender, face gender, or version (all \u003cem\u003ep\u003c/em\u003e \u0026gt; .10). Model diagnostics using the DHARMa package indicated no signs of overdispersion (\u003cem\u003ep =\u0026nbsp;\u003c/em\u003e.984) or zero-inflation (\u003cem\u003ep =\u0026nbsp;\u003c/em\u003e.938), supporting the adequacy of model fit. Tukey-adjusted post-hoc comparisons showed that SHR for affiliative and dominance smiles were significantly lower than for neutral, reward, and disgust expressions (all \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Additionally, affiliative expressions yielded significantly higher SHR than dominance expressions (\u003cem\u003ep =\u0026nbsp;\u003c/em\u003e.075, marginal).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eUnbiased Hit Rate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccuracy estimates are affected by two kinds of biases - the relative number of items for each emotion category presented (presentation bias) and the relative utilization of different emotion categories by the participants (response bias) (B\u0026auml;nziger et al., 2012). UHR is an accuracy index, which considers both biases by calculating \u0026ldquo;the joint probability that a stimulus is correctly identified (given that it is presented) and that a response is correctly used (given that it is used)\u0026rdquo; (Wagner, 1993). Here we calculated the UHR following Bijsterbosch et al. (2021) in two steps: First, a choice matrix was created with chosen emotion expressions as columns and intended emotional expressions as rows. Second, the number of ratings in each cell was squared and divided by the product of the marginal values of the corresponding row and column (Note: there were no missing data). Since UHR is a proportion, the values were arcsine-transformed to correct for skewed variances of decimal fractions obtained from count (Bishara \u0026amp; Hittner, 2012; Goeleven et al., 2008). The transformed scores can range from 0 to 1.57, the arcsine of 1, representing perfect identification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDescriptive statistics for arcsine-transformed UHRs followed a similar pattern with SHRs (summarized in Table 2). Neutral, reward, and disgust expressions achieved high UHRs (e.g., arcsine \u003cem\u003eM =\u003c/em\u003e 1.42, 1.37, and 1.39 respectively), whereas affiliative and dominance yielded lower performance (arcsine \u003cem\u003eM =\u003c/em\u003e 1.00 and 1.03). Similarly, unbiased hit rates (UHR) exceeded chance levels across all expressions, further confirming the recognizability of the synthesized emotional stimuli.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo compare the UHRs across emotion categories, we applied a linear mixed effects model (LMM) on arcsine-transformed unbiased UHRs, treating expression category, group, rater gender, model gender, and mirror version as fixed effects, and including a random intercept for each rater (CASE). The model converged with a REML criterion of 732.7. Relative to the reference category (neutral), expression type showed significant effects. Specifically, compared to neutral, the affiliative condition yielded a significantly lower UHR (\u003cem\u003e\u0026beta;\u003c/em\u003e = -0.425, \u003cem\u003eSE\u003c/em\u003e = 0.026, \u003cem\u003et\u003c/em\u003e = -16.50, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001), as did the dominance condition (\u003cem\u003e\u0026beta;\u003c/em\u003e = -0.394, \u003cem\u003eSE\u003c/em\u003e = 0.026, \u003cem\u003et\u003c/em\u003e = -15.26, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001). In addition, reward exhibited a small yet significant effect (\u003cem\u003e\u0026beta;\u003c/em\u003e = -0.052, \u003cem\u003eSE\u003c/em\u003e = 0.026, \u003cem\u003et\u003c/em\u003e = -2.02, \u003cem\u003ep\u003c/em\u003e = 0.043), whereas disgust did not differ from neutral (\u003cem\u003e\u0026beta;\u003c/em\u003e = 0.012, \u003cem\u003eSE\u003c/em\u003e = 0.026, \u003cem\u003et\u003c/em\u003e = 0.48, \u003cem\u003ep\u003c/em\u003e = 0.630). Other factors (face gender, rater gender, version, and group) did not significantly affect UHR (all \u003cem\u003ep\u003c/em\u003e \u0026gt; 0.10).\u003c/p\u003e\n\u003cp\u003ePost-hoc pairwise comparisons (Tukey-adjusted) revealed that neutral differed significantly from both affiliative (\u003cem\u003eestimate\u003c/em\u003e = 0.425, \u003cem\u003et\u003c/em\u003e = 16.50, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001) and dominance (\u003cem\u003eestimate\u003c/em\u003e = 0.394, \u003cem\u003et\u003c/em\u003e = 15.26, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001), whereas differences between neutral and either reward (\u003cem\u003eestimate\u003c/em\u003e = 0.052, \u003cem\u003et\u003c/em\u003e = 2.02, \u003cem\u003ep\u003c/em\u003e = 0.2553) or disgust (\u003cem\u003eestimate\u003c/em\u003e = -0.012, \u003cem\u003et\u003c/em\u003e = -0.48, \u003cem\u003ep\u003c/em\u003e = 0.9890) were not significant. Additionally, affiliative differed significantly from both disgust (\u003cem\u003eestimate\u003c/em\u003e = -0.437, \u003cem\u003et\u003c/em\u003e = -16.98, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001) and reward (\u003cem\u003eestimate\u003c/em\u003e = -0.373, \u003cem\u003et\u003c/em\u003e = -14.47, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001), while the difference between affiliative and dominance was non-significant (\u003cem\u003ep\u003c/em\u003e = 0.7413). Further comparisons (e.g., disgust vs. dominance: \u003cem\u003eestimate\u003c/em\u003e = 0.406, \u003cem\u003et\u003c/em\u003e = 15.74, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001; dominance vs. reward: \u003cem\u003eestimate\u003c/em\u003e = -0.341, \u003cem\u003et\u003c/em\u003e = -13.24, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001) underscore the robust impact of expression type on UHR.\u003c/p\u003e\n\u003cp\u003eOverall, our findings on UHR aligns with SHR: neutral, disgust, and reward conditions yield higher recognition performance than affiliative and dominance, while group assignment, rater gender, model gender, and version have no systematic impact.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConfusion Matrix\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccording to B\u0026auml;nziger et al. (2012) a Confusion Matrix is an appropriate and complete way to provide all \u0026nbsp;the relevant information about potential bias since different attempts to propose indices that are corrected for chance guessing and/or response bias (e.g., the UHR suggested by \u0026nbsp;Wagner, 1993) generally mask the origin of the potential biases and do not allow to take these \u0026nbsp;into account in interpreting the results. Thus, we also present the confusion matrix (see Fig. 3) representing the classification results of emotional facial expressions, illustrating the percentage of chosen emotional categories (e.g., reward, disgust, affiliative, dominance, neutral, other) for each intended emotional expression. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eA close look at the confusion matrix reveals several notable patterns in the participants\u0026apos; responses to different facial expressions. Faces with intended affiliative smiles were frequently confused with neutral expressions (7.5%) and reward smiles (17.7%), suggesting significant perceptual overlap between these expressions. Similarly, intended dominance expressions were misclassified as affiliative (10.9%) disgust (7.8%) or others (9.8%), highlighting some confusion in identifying dominance-related expressions. Reward expressions were correctly identified in most cases, though there was some misclassification as affiliative (5.9%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eArousal and plausibility\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn our study, the mean scores of the arousal ratings were 2.63 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.66) for neutral expressions, 4.50 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.60) for affiliative, 5.44 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.66) for disgust, 5.29 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.59) for dominance, and 6.72 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.39) for reward. Mean scores for plausibility ratings were 5.21 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.26) for neutral expressions, 4.66 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.39) for affiliative, 4.23 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.41) for disgust, 4.06 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.43) for dominance, and 5.02 (\u003cem\u003eSD =\u0026nbsp;\u003c/em\u003e1.33) for reward.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo compare the arousal and plausibility scores across expression categories, linear mixed‐effects models (LMM) were fitted to these scores with fixed effects of expression category, group, rater gender, model gender, and image version, and a random intercept for CASE. Models were fitted using restricted maximum likelihood (REML). The intercept was estimated at 2.759 (\u003cem\u003eSE =\u003c/em\u003e 0.223, \u003cem\u003et\u003c/em\u003e(\u0026asymp;123.9) = 12.36, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e0.001), representing the predicted arousal score for the reference level (neutral) of expression category. Compared with neutral expressions, the fixed effects estimates for expression category were as follows: affiliative: 1.865 (\u003cem\u003eSE =\u003c/em\u003e 0.0408, \u003cem\u003et\u003c/em\u003e = 45.73, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001); disgust: 2.808 (\u003cem\u003eSE =\u003c/em\u003e 0.0408, \u003cem\u003et\u003c/em\u003e = 68.88, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001); dominance: 2.661 (\u003cem\u003eSE =\u003c/em\u003e 0.0408, \u003cem\u003et\u003c/em\u003e = 65.26, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001); and reward: 4.088 (\u003cem\u003eSE =\u003c/em\u003e 0.0408, \u003cem\u003et\u003c/em\u003e = 100.26, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001). Tukey‐adjusted pairwise comparisons confirmed that ratings in the neutral condition were significantly lower than each other expression (all \u003cem\u003ep\u003c/em\u003e-values \u003cem\u003e\u0026lt;\u0026nbsp;\u003c/em\u003e0.001). In comparisons between non-neutral categories, all differences were significant (\u003cem\u003ep \u0026lt;\u003c/em\u003e 0.001), with the exception of the disgust\u0026ndash;dominance contrast, which, though smaller, remained statistically significant (\u003cem\u003ep =\u003c/em\u003e 0.0028). Additionally, the effect of model gender was significant: images showing male faces received lower arousal ratings than those showing female faces (coefficient = \u0026ndash;0.145, \u003cem\u003eSE =\u003c/em\u003e 0.0365, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e0.001);\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eA similar LMM was applied to the plausibility scores. The intercept was estimated at 4.833 (\u003cem\u003eSE =\u003c/em\u003e 0.190, \u003cem\u003et\u003c/em\u003e(\u0026asymp;124) = 25.46, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e0.001). Compared with neutral expressions, the fixed effects estimates were \u0026ndash;0.553 (\u003cem\u003eSE =\u003c/em\u003e 0.035, \u003cem\u003et\u003c/em\u003e = \u0026ndash;15.79, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001) for affiliative, \u0026ndash;0.986 (\u003cem\u003eSE =\u003c/em\u003e 0.035, \u003cem\u003et\u003c/em\u003e = \u0026ndash;28.15, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001) for disgust, \u0026ndash;1.150 (\u003cem\u003eSE =\u003c/em\u003e 0.035, \u003cem\u003et\u003c/em\u003e = \u0026ndash;32.84, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001) for dominance, and \u0026ndash;0.197 (\u003cem\u003eSE =\u003c/em\u003e 0.035, \u003cem\u003et\u003c/em\u003e = \u0026ndash;5.61, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e 0.001) for reward. Tukey‐adjusted comparisons indicated that all pairwise comparisons between expression categories were statistically significant (\u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e0.001). Moreover, the effect of model gender was significant, with images of male faces receiving higher plausibility ratings than female faces (coefficient = 0.248, \u003cem\u003eSE =\u003c/em\u003e 0.0313, \u003cem\u003ep \u0026lt;\u0026nbsp;\u003c/em\u003e0.001). Additionally, stimulus version (original vs. mirrored) significantly affected plausibility ratings, with R-versions (mirrored images) rated slightly but significantly less plausible than L-versions (original orientation) (coefficient = \u0026ndash;0.169, \u003cem\u003eSE\u003c/em\u003e = 0.044, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001).\u003c/p\u003e\n\u003cp\u003eOverall, although the Shapiro\u0026ndash;Wilk and Levene tests for the residuals were statistically significant\u0026mdash;likely due to the large sample sizes\u0026mdash;visual inspection of residual plots suggested that deviations from normality and homogeneity of variance were minor. Overall, both LMM models revealed statistically significant differences in ratings across expression category levels, and the potential effects of model gender in shaping perceptual evaluations of the images.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInterrater reliability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess the reliability of the ratings across participants for arousal, plausibility, and emotion categorization, we calculated Fleiss\u0026apos; Kappa (for categorical data) and Intraclass Correlation Coefficients (ICCs) (for continuous data). Fleiss\u0026rsquo; Kappa estimates the agreement among multiple raters for nominal categories (Fleiss, 1971), while ICCs provide a measure of consistency in ratings across raters (Koo \u0026amp; Li, 2016; Shrout \u0026amp; Fleiss, 1979).A linear mixed-effects model with Group (i.e., each of the 10 rater lists) as a fixed effect was first used to examine potential between-group differences. Since no significant group effects were found for any of the rating types, we report mean Kappa and ICC values collapsed across all 10 groups.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFleiss\u0026apos; Kappa.\u0026nbsp;\u003c/strong\u003eThe agreement between participants in categorizing emotional expressions was substantial, with an overall Fleiss\u0026rsquo; Kappa of 0.65, indicating reliable identification of the intended expression categories.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eICC.\u0026nbsp;\u003c/strong\u003eTo objectively select the appropriate ICCs, the model fit (deviance information criterion) was calculated for all possible ICC models. The model with the best fit to our data was ICC 2, applying a two-way cross-classified multilevel model. The reliability of the judges\u0026apos; ratings for arousal and plausibility was assessed using ICC (2,1) and ICC (2,\u003cem\u003e\u0026nbsp;k\u003c/em\u003e). ICC (2, 1) intraclass correlation coefficient, each image is measured by each rater with reliability estimated from a single measurement; ICC (2, \u003cem\u003ek\u003c/em\u003e) as before, but here reliability is calculated by imputing the average of the k raters\u0026rsquo; measurements (Bijsterbosch et al., 2021).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor arousal, the mean ICC (2,1) was 0.44, indicating moderate agreement between individual raters, while the ICC(2,k) was 0.97, reflecting excellent reliability at the group level. For plausibility, ICC (2,1) was lower at 0.14, indicating poor reliability across individuals, but the ICC(2,k) reached 0.88, suggesting good average agreement across raters (based on standard guidelines from Koo \u0026amp; Li, 2016).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRanking of Expressors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn order to assess the relative quality of the expressions derived for the different individuals on which the expressions were based we ranked the performance of all expressors. At first we calculated the UHR and plausibility scores across all emotion categories per expressor. As described above, the UHR values had been arcsine-transformed to stabilize variance. A composite score was then derived by averaging the Z-scores of the arcsine-transformed UHRs and the plausibility scores. Our 90 expressors were ranked based on this composite score, grouped into ten rank-based categories, with gender distributions illustrated through pie charts (see Fig. 4). See Appendix for an overview of all individual expressors.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eExtending prior applications of FACSGen, the present study introduced a pipeline for synthesizing socially nuanced emotional expressions from neutral portraits of real individuals. We constructed a rigorously validated stimulus set of 900 images, with careful control of visual confounds. Recognition accuracy was high overall\u0026mdash;particularly for reward and basic expressions\u0026mdash;while affiliative and dominance smiles were more frequently confused with related categories, reflecting their subtle and socially overlapping nature. Arousal and plausibility ratings also varied systematically by expression type and model gender. These findings demonstrate not only the technical feasibility of AU-based expression synthesis but also its potential to advance research on subtle, socially meaningful facial expressions.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eEmotion recognition accuracy\u003c/h2\u003e \u003cp\u003eRecognition accuracy is a crucial measure of validity for emotional stimulus sets. In our validation study, we systematically evaluated how well emotional expressions synthesized from neutral images of real expressors\u0026mdash;particularly subtle variants of smiles\u0026mdash;can be identified. To this aim, we employed Simple Hit Rates (SHR) and Unbiased Hit Rates (UHR) to evaluate overall recognition accuracy, and visualized response tendencies using a confusion matrix illustrating perceptual overlaps among certain expressions.\u003c/p\u003e \u003cp\u003eThe mean SHR for neutral expressions was remarkably high (92.6%), surpassing previously reported hit rates, such as the 87% reported by Ebner et al., (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). This likely reflects the fact that many of the neutral-expression portraits used here were sourced from validated datasets and had already undergone a selection process based on recognizability. The demographic profile of our participants\u0026mdash;exclusively younger adults\u0026mdash;may have also contributed, as younger individuals typically perform better in facial emotion recognition than older adults.\u003c/p\u003e \u003cp\u003eAmong emotional expressions, disgust demonstrated high recognition accuracy (91.5%), exceeding previously reported benchmarks such as the 68.6% accuracy reported by Krumhuber et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) during the initial validation of FACSGen 2.0-generated disgust expressions. The 'happiness' category used in previous studies closely aligns with the reward smiles in our study. The mean recognition accuracy for reward smiles in the present study (92.1%) was somewhat higher than the happiness recognition accuracy reported by Krumhuber et al., (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) (88.46%). These findings highlight the efficacy of using carefully controlled AU configurations in FACSGen 2.0 to generate highly distinguishable emotional stimuli.\u003c/p\u003e \u003cp\u003eNot surprisingly, recognition performance differed markedly from reward smiles for more subtle variants of smiles; SHR for affiliative and dominance smiles was at 69.2% and 66.0%, respectively. The confusion matrix revealed significant perceptual overlaps among these subtle expressions: affiliative smiles were frequently misclassified as reward smiles (17.7%) or neutral expressions (7.5%), while dominance smiles were commonly confused with affiliative smiles (10.9%), disgust (7.8%), or categorized as \"other\" (9.8%). These misclassification patterns suggest inherent perceptual similarities likely due to overlapping Action Unit (AU) configurations, aligning with previous findings by Rychlowska et al. (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) indicating that subtle variations in smiles pose perceptual challenges.\u003c/p\u003e \u003cp\u003eThe observations from SHR and the confusion matrix were supported by the Unbiased Hit Rate analyses, which account for both presentation and response biases. Neutral expressions, disgust, and reward expressions consistently yielded higher recognition performance compared to affiliative and dominance smiles, though importantly, accuracy remained far above chance levels for all emotion categories employed here.\u003c/p\u003e \u003cp\u003eIn summary, our findings confirm the effectiveness of predefined AU configurations in producing distinguishable emotional expressions via FACSGen 2.0. At the same time, they emphasize specific perceptual challenges associated with subtle variants of smiles, reinforcing previous research and highlighting important areas for future methodological refinement.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eArousal and plausibility\u003c/h2\u003e \u003cp\u003eOur analyses of arousal ratings revealed clear distinctions among expression categories that align with expectations. Consistent with previous research (e.g., Goeleven et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2008\u003c/span\u003e), neutral expressions elicited the lowest arousal ratings for the image (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;2.63), serving as an effective baseline for other emotional expressions. In contrast, disgust expressions produced substantially higher arousal ratings (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;5.44), reflecting their established role in signalling negative affective states (Rozin et al., \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e1994\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eNotably, our primary interest concerned the subtle variants of smiles. Reward smiles received the highest arousal ratings (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;6.72), which is consistent with their communicative function of conveying positive emotional intensity and social engagement (Sauter, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Affiliative (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;4.50) and dominance smiles (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;5.29) elicited intermediate arousal levels, consistent with their more nuanced communicative roles. Affiliative smiles, aiming primarily at signalling warmth and cooperation, understandably evoke more moderate arousal. Dominance smiles, signalling social status or assertiveness, produced higher arousal than affiliative but lower than reward smiles, reflecting their social signalling function as documented in prior studies (Rychlowska et al., \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003ePlausibility ratings about how likely it was that the synthetic images were derived from pictures of real people were conducted to assess the apparent naturalness of the stimuli, which is very important for their usefulness in experimental studies (Diel et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The plausibility ratings indicated distinctions among the expressions. Neutral expressions were perceived as most plausibly human-based (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;5.21), given their minimal modifications and frequent occurrence in natural interactions. Synthetic reward smiles (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;5.02) obtained high plausibility ratings comparable to neutral faces, consistent with findings that smiles reflecting genuine positive affect tend to be perceived as natural and authentic (Ekman \u0026amp; Friesen, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1982\u003c/span\u003e). In contrast, disgust (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;4.23) and dominance smiles (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;4.06) received lower plausibility scores, possibly due to their reliance on less frequently encountered Action Units (e.g., AU9 and AU10), making them less common in everyday facial displays and thus perceived as less genuine or natural (Dyck et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). The plausibility of affiliative smiles (\u003cem\u003eM\u0026thinsp;=\u003c/em\u003e\u0026thinsp;4.66) fell in-between, suggesting that subtler smile expressions, while socially meaningful, might challenge observers\u0026rsquo; perceptions of naturalness due to their ambiguity or less frequent occurrence (Rychlowska et al., \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eBeyond expression type, we also observed systematic differences based on the model's gender, despite identical AU parameters across expressors. Specifically, female expressors elicited significantly higher arousal ratings compared to male expressors, aligning with previous research suggesting observers perceive female emotional expressions as more affectively intense or expressive (Hess et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2000\u003c/span\u003e; Plant et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2000\u003c/span\u003e). Conversely, male expressors received higher plausibility ratings, indicating their emotional expressions were perceived as more genuine or believable, though relatively unnatural hairstyles assigned to female expressors might partially explain this finding. Future studies could address this by concealing external facial features with oval masks or using more realistic hairstyles. Alternatively, societal stereotypes or expectations may also contribute to this result, as emotional displays by men could be seen as more intentional or credible, whereas heightened expressivity in women might sometimes be interpreted as exaggerated or performative (Kring \u0026amp; Gordon, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1998\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOverall, these findings confirm that our synthesis method successfully generates expressions that vary meaningfully along both arousal and plausibility dimensions. The clear differentiation between reward, affiliative, and dominance smiles highlights the perceptual relevance of subtle AU variations in conveying affective meaning. Moreover, the observed gender differences in perceived intensity and naturalness, despite identical AU settings, suggest that visual appearance and social expectations interact in shaping emotion perception. Together, these results underscore the importance of considering both low-level visual cues and higher-order social factors in the design and validation of emotional facial stimuli.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eInterrater reliability\u003c/h2\u003e \u003cp\u003eOur analysis of interrater reliability indicates a generally robust consistency among participants. The overall Fleiss' Kappa value of 0.65 for emotion categorization demonstrates substantial agreement (Landis \u0026amp; Koch, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e1977\u003c/span\u003e), suggesting that participants were consistent in identifying the intended expressions provided in the current study. This level of agreement supports the validity of our categorization process; for arousal ratings, the single-rater reliability, as measured by ICC (2,1), was moderate at 0.44, while the average reliability across participants, ICC (2, \u003cem\u003ek\u003c/em\u003e), reached an excellent level of 0.97. This disparity highlights that, although individual ratings might vary, the aggregate mean of multiple participants yields highly reliable arousal estimates; In the case of plausibility ratings, a similar pattern emerged: the single-rater ICC (2,1) was low (0.14), indicating considerable variability at the individual level. However, the aggregate reliability (ICC (2, \u003cem\u003ek\u003c/em\u003e)) was strong at 0.88. These results underscore the importance of aggregating ratings in studies of this kind, as group averages effectively mitigate individual differences and enhance overall reliability.\u003c/p\u003e \u003cp\u003eCollectively, these reliability indices confirm that our methodological approach yields consistent ratings across multiple evaluators, thereby reinforcing the robustness of our experimental design and the validity of our subsequent analyses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eRanking of Expressors\u003c/h2\u003e \u003cp\u003eAdditionally, to further evaluate and compare the overall quality of synthesized stimuli, we developed a composite ranking score for each expressor by integrating standardized unbiased hit rates (UHR) and plausibility ratings. This composite measure enabled us to systematically identify which expressors performed optimally across multiple evaluation criteria, providing practical guidance for researchers selecting the most effective emotional stimuli for experimental purposes. The ranking also highlighted potential differences related to model gender distribution with images derived from male expressors scoring higher than female expressors, which could inform future studies aimed at balancing stimuli to mitigate biases or investigate gender-specific effects in emotion perception. We speculate that the generally smoother appearance of female faces makes it harder to effectively apply the AU manipulation by the software employed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003ePerspectives and Limitations\u003c/h2\u003e \u003cp\u003eThe current procedure offers researchers considerable flexibility, allowing image selection based on various parameters\u0026mdash;such as model gender, hemiface, expression category, recognition accuracy, arousal, and plausibility\u0026mdash;to meet diverse research needs. Naturally, the pipeline can be extended to generate other or compound emotions, provided that the intended emotional expressions can be precisely defined in terms of Action Units (AUs). Moreover, our open-source code\u003csup\u003e2\u003c/sup\u003e base enables researchers to independently validate their own stimuli.\u003c/p\u003e \u003cp\u003eAnother promising avenue for future research is the inclusion of dynamic facial expressions, which can be readily created using FACSGen 2.0. Prior research has shown that dynamic expressions enhance affect recognition, increase perceived authenticity, and aid in distinguishing genuine from posed expression (Krumhuber et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Additionally, expanding the database to include faces from more diverse racial backgrounds is both feasible and beneficial. Because our method operates on neutral base portraits, which are readily available or easily obtained, extending this framework requires no specialized acting skills or training on the part of expressors.\u003c/p\u003e \u003cp\u003eSeveral limitations of the present work should be acknowledged. First, because our method is based on neutral portraits sourced from different existing databases, any stimuli originating from external databases are available only by contacting the respective database owners due to redistribution restrictions (see Appendix). However, since the validation analysis code is open-source, researchers can easily apply the synthesis pipeline and validation procedures to their own image sources. Additionally, experimental stimuli generated from our in-house database are directly available upon request, facilitating adaptation or extension to meet diverse research objectives.\u003c/p\u003e \u003cp\u003eSecond, although mirrored (R-version) and original (L-version) images did not differ in recognition accuracy or arousal, plausibility ratings were slightly lower for mirrored versions, especially for dominance and reward smiles. To objectively acknowledge and accurately report this observed difference, we decided to treat the two versions separately instead of merging the mirrored and original versions as initially intended. Consequently, the number of raters per expression per individual was reduced from 26 to 13. While this sample size per image aligns with typical standards in facial-expression validation studies (Delicato, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), it nonetheless raises potential concerns regarding the stability and generalizability of ratings at the item level. Future research involving mirrored versions or subtle perceptual differences should therefore consider recruiting a greater number of participants per image version to enhance the robustness and representativeness of individual stimulus ratings. The observed plausibility difference likely reflects subtle lighting asymmetries combined with observers' inherent preference for the left side of the face (Lindell, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Powell \u0026amp; Schirillo, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Future research might explore under what conditions such lateralization effects significantly influence plausibility without affecting recognition accuracy or perceived arousal.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis study presents a cost-effective pipeline for generating high-quality emotional stimuli from neutral facial images while preserving individual identity. The synthesized expressions\u0026mdash;reward, affiliative, and dominance smiles, disgust, and neutral\u0026mdash;were validated across emotional content, arousal, and plausibility. Notably, the confusions between affiliative and dominance smiles underscore the interpretive complexity of socially nuanced expressions. Together, these findings highlight the challenges of perceiving subtle affective signals and offer a validated, scalable tool to support future research on facial emotion recognition and social communication.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by the China Scholarship Council under Grant No. 202208210077.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of interest/Competing interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing financial or non-financial interests related to this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study protocol was approved by the Ethics Committee of the Department of Psychology, Humboldt-Universit\u0026auml;t zu Berlin (Approval No. 2024-17). All procedures complied with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll participants provided informed consent electronically before beginning the online survey.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eParticipants provided informed consent for the use and publication of their anonymized rating data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was preregistered on OSF. The preregistration and all anonymized rating data are available at https://osf.io/j4dfr/?view_only=8f6a7b25d26f4b79a6764eb398594285 (Open-Ended Registration, registered on August 7, 2024). The full set of in-house-derived stimuli is available upon request from the corresponding author. Any materials originating from external databases remain available via their original owners due to redistribution restrictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eJin Gao: Conceptualization; data curation; formal analysis; investigation; methodology; software; visualization; writing \u0026ndash; original draft; writing \u0026ndash; review and editing.\u003c/li\u003e\n \u003cli\u003eWerner Sommer (Prof.): Conceptualization; funding acquisition; methodology; project administration; resources; supervision; writing \u0026ndash; review and editing.\u003c/li\u003e\n \u003cli\u003eRasha Abdel Rahman (Prof.): Conceptualization; funding acquisition; supervision; writing \u0026ndash; review and editing.\u003c/li\u003e\n \u003cli\u003eWei-Jun Li (Prof.): Supervision; writing \u0026ndash; review and editing.\u003c/li\u003e\n\u003c/ul\u003e\u003cp\u003e\u003cstrong\u003ePublic Significance Statements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot every smile signals happiness\u0026mdash;some build rapport, others assert status. We offer a rigorous yet easy-to-use pipeline, built with specialized face-editing software, that turns neutral photos into controlled sets of reward, affiliative, and dominance smiles and validates them for recognition accuracy, arousal, and plausibility, giving researchers a clear recipe for studying nuanced emotion signals. The results reveal shared cues but frequent confusions among these smiles, reminding the public that even tiny facial shifts shape what we see and convey.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eB\u0026auml;nziger, T., Mortillaro, M., \u0026amp; Scherer, K. R. (2012). Introducing the Geneva Multimodal Expression Corpus for experimental research on emotion perception. \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(5), 1161\u0026ndash;1179. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0025827\u003c/span\u003e\u003cspan address=\"10.1037/a0025827\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBehrouzi, T., Shahroudnejad, A., \u0026amp; Mousavi, P. (2025). Maskrenderer: 3D-infused multi-mask realistic face reenactment. \u003cem\u003ePattern Recognition\u003c/em\u003e, \u003cem\u003e157\u003c/em\u003e, 110891. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.patcog.2024.110891\u003c/span\u003e\u003cspan address=\"10.1016/j.patcog.2024.110891\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBijsterbosch, G., Mobach, L., Verpaalen, I. A. M., Bijlstra, G., Hudson, J. L., Rinck, M., \u0026amp; Klein, A. M. (2021). Validation of the child models of the Radboud Faces Database by children. \u003cem\u003eInternational Journal of Behavioral Development\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e(2), 146\u0026ndash;152. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0165025420935631\u003c/span\u003e\u003cspan address=\"10.1177/0165025420935631\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBishara, A. J., \u0026amp; Hittner, J. B. (2012). Testing the significance of a correlation with nonnormal data: Comparison of Pearson, Spearman, transformation, and resampling approaches. \u003cem\u003ePsychological Methods\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(3), 399\u0026ndash;417. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0028087\u003c/span\u003e\u003cspan address=\"10.1037/a0028087\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorod, J. C., Haywood, C. S., \u0026amp; Koff, E. (1997). Neuropsychological aspects of facial asymmetry during emotional expression: A review of the normal adult literature. \u003cem\u003eNeuropsychology Review\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1), 41\u0026ndash;60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF02876972\u003c/span\u003e\u003cspan address=\"10.1007/BF02876972\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBusin, Y., Lukasova, K., Asthana, M. K., \u0026amp; Macedo, E. C. (2018). Hemiface differences in visual exploration patterns when judging the authenticity of facial expressions. \u003cem\u003eFrontiers in Psychology\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 2332. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fpsyg.2017.02332\u003c/span\u003e\u003cspan address=\"10.3389/fpsyg.2017.02332\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCrookes, K., Ewing, L., Gildenhuys, J., Kloth, N., Hayward, W. G., Oxner, M., Pond, S., \u0026amp; Rhodes, G. (2015). How well do computer-generated faces tap face expertise? \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(11), e0141353. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0141353\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0141353\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDawel, A., Miller, E. J., Horsburgh, A., \u0026amp; Ford, P. (2022). A systematic survey of face stimuli used in psychological research 2000\u0026ndash;2020. \u003cem\u003eBehavior Research Methods\u003c/em\u003e, \u003cem\u003e54\u003c/em\u003e(4), 1889\u0026ndash;1901. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/s13428-021-01705-3\u003c/span\u003e\u003cspan address=\"10.3758/s13428-021-01705-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeBruine, L., \u0026amp; Jones, B. (2017). Face Research Lab London Set [Dataset]. \u003cem\u003eFigshare\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.6084/m9.figshare.5047666.v1\u003c/span\u003e\u003cspan address=\"10.6084/m9.figshare.5047666.v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDelicato, L. S. (2020). A robust method for measuring an individual\u0026rsquo;s sensitivity to facial expressions. \u003cem\u003eAttention Perception \u0026amp; Psychophysics\u003c/em\u003e, \u003cem\u003e82\u003c/em\u003e(6), 2924\u0026ndash;2936. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/s13414-020-02043-w\u003c/span\u003e\u003cspan address=\"10.3758/s13414-020-02043-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDiel, A., Weigelt, S., \u0026amp; Macdorman, K. F. (2021). A meta-analysis of the uncanny valley\u0026rsquo;s independent and dependent variables. \u003cem\u003eJournal of Human-Robot Interaction\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003eArticle 1. https://doi.org/10.1145/3470742\u003c/span\u003e\u003cspan address=\"Article 1. 10.1145/3470742\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDyck, M., Winbeck, M., Leiberg, S., Chen, Y., Gur, R. C., \u0026amp; Mathiak, K. (2008). Recognition profile of emotions in natural and virtual faces. \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(11), e3628. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0003628\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0003628\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEbner, N. C., Riediger, M., \u0026amp; Lindenberger, U. (2010). FACES\u0026mdash;A database of facial expressions in young, middle-aged, and older women and men: Development and validation. \u003cem\u003eBehavior Research Methods\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e(1), 351\u0026ndash;362. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/BRM.42.1.351\u003c/span\u003e\u003cspan address=\"10.3758/BRM.42.1.351\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkman, P. (1976). \u003cem\u003ePictures of facial affect\u003c/em\u003e. Consulting Psychologists.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkman, P., \u0026amp; Friesen, W. V. (1982). Felt, false, and miserable smiles. \u003cem\u003eJournal of Nonverbal Behavior\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(4), 238\u0026ndash;252. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF00987191\u003c/span\u003e\u003cspan address=\"10.1007/BF00987191\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkman, P., \u0026amp; Rosenberg, E. L. (1997). \u003cem\u003eWhat the Face Reveals: Basic and Applied Studies of Spontaneous Expression Using the Facial Action Coding System (FACS)\u003c/em\u003e. Oxford University Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFleiss, J. L. (1971). Measuring nominal scale agreement among many raters. \u003cem\u003ePsychological Bulletin\u003c/em\u003e, \u003cem\u003e76\u003c/em\u003e(5), 378\u0026ndash;382. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/h0031619\u003c/span\u003e\u003cspan address=\"10.1037/h0031619\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoeleven, E., De Raedt, R., Leyman, L., \u0026amp; Verschuere, B. (2008). The Karolinska Directed Emotional Faces: A validation study. \u003cem\u003eCognition and Emotion\u003c/em\u003e, \u003cem\u003e22\u003c/em\u003e(6), 1094\u0026ndash;1118. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/02699930701626582\u003c/span\u003e\u003cspan address=\"10.1080/02699930701626582\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHareli, S., \u0026amp; Hess, U. (2012). The social signal value of emotions. \u003cem\u003eCognition and Emotion\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(3), 385\u0026ndash;389. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/02699931.2012.665029\u003c/span\u003e\u003cspan address=\"10.1080/02699931.2012.665029\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHess, U., Sen\u0026eacute;cal, S., Kirouac, G., Herrera, P., Philippot, P., \u0026amp; Kleck, R. E. (2000). Emotional expressivity in men and women: Stereotypes and self-perceptions. \u003cem\u003eCognition and Emotion\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(5), 609\u0026ndash;642. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/02699930050117648\u003c/span\u003e\u003cspan address=\"10.1080/02699930050117648\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHorstmann, G. (2003). What do facial expressions convey: Feeling states, behavioral intentions, or action requests? \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(2), 150\u0026ndash;166. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/1528-3542.3.2.150\u003c/span\u003e\u003cspan address=\"10.1037/1528-3542.3.2.150\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKammoun, A., Slama, R., Tabia, H., Ouni, T., \u0026amp; Abid, M. (2022). Generative adversarial networks for face generation: A survey. \u003cem\u003eACM Computing Surveys\u003c/em\u003e, \u003cem\u003e55\u003c/em\u003e(5), 94:1\u0026ndash;9437. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3527850\u003c/span\u003e\u003cspan address=\"10.1145/3527850\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaulard, K., Cunningham, D. W., B\u0026uuml;lthoff, H. H., \u0026amp; Wallraven, C. (2012). The MPI facial expression database\u0026mdash;A validated database of emotional and conversational facial expressions. \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(3), e32321. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0032321\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0032321\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlingner, C. M., \u0026amp; Guntinas-Lichius, O. (2023). Facial expression and emotion. \u003cem\u003eLaryngo-Rhino-Otologie\u003c/em\u003e, \u003cem\u003e102\u003c/em\u003e(Suppl 1), S115\u0026ndash;S125. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1055/a-2003-5687\u003c/span\u003e\u003cspan address=\"10.1055/a-2003-5687\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoo, T. K., \u0026amp; Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. \u003cem\u003eJournal of Chiropractic Medicine\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 155\u0026ndash;163. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jcm.2016.02.012\u003c/span\u003e\u003cspan address=\"10.1016/j.jcm.2016.02.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKring, A. M., \u0026amp; Gordon, A. H. (1998). Sex differences in emotion: Expression, experience, and physiology. \u003cem\u003eJournal of Personality and Social Psychology\u003c/em\u003e, \u003cem\u003e74\u003c/em\u003e(3), 686\u0026ndash;703. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0022-3514.74.3.686\u003c/span\u003e\u003cspan address=\"10.1037/0022-3514.74.3.686\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrumhuber, E. G., Kappas, A., \u0026amp; Manstead, A. S. R. (2013). Effects of dynamic aspects of facial expressions: A review. \u003cem\u003eEmotion Review\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(1), 41\u0026ndash;46. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/1754073912451349\u003c/span\u003e\u003cspan address=\"10.1177/1754073912451349\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrumhuber, E. G., Tamarit, L., Roesch, E. B., \u0026amp; Scherer, K. R. (2012). FACSGen 2.0 animation software: Generating three-dimensional FACS-valid facial expressions for emotion research. \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(2), 351\u0026ndash;363. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0026632\u003c/span\u003e\u003cspan address=\"10.1037/a0026632\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLandis, J. R., \u0026amp; Koch, G. G. (1977). The measurement of observer agreement for categorical data. \u003cem\u003eBiometrics\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(1), 159\u0026ndash;174. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2307/2529310\u003c/span\u003e\u003cspan address=\"10.2307/2529310\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLang, P. J. (1980). Behavioral treatment and bio-behavioral assessment: Computer applications. In J. B. Sidowski, J. H. Johnson, \u0026amp; T. A. Williams (Eds.), \u003cem\u003eTechnology in mental health care delivery systems\u003c/em\u003e (pp. 119\u0026ndash;137). Ablex.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLindell, A. K. (2013). The silent social/emotional signals in left and right cheek poses: A literature review. \u003cem\u003eLaterality\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(5), 612\u0026ndash;624. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/1357650X.2012.737330\u003c/span\u003e\u003cspan address=\"10.1080/1357650X.2012.737330\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa, D. S., Correll, J., \u0026amp; Wittenbrink, B. (2015). The Chicago face database: A free stimulus set of faces and norming data. \u003cem\u003eBehavior Research Methods\u003c/em\u003e, \u003cem\u003e47\u003c/em\u003e(4), 1122\u0026ndash;1135. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/s13428-014-0532-5\u003c/span\u003e\u003cspan address=\"10.3758/s13428-014-0532-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartin, J., Rychlowska, M., Wood, A., \u0026amp; Niedenthal, P. M. (2017). Smiles as multipurpose social signals. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e21\u003c/em\u003e(11), 864\u0026ndash;877. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.tics.2017.08.007\u003c/span\u003e\u003cspan address=\"10.1016/j.tics.2017.08.007\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehu, M., Mortillaro, M., B\u0026auml;nziger, T., \u0026amp; Scherer, K. R. (2012). Reliable facial muscle activation enhances recognizability and credibility of emotional expression. \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(4), 701\u0026ndash;715. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0026632\u003c/span\u003e\u003cspan address=\"10.1037/a0026632\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollahosseini, A., Hasani, B., \u0026amp; Mahoor, M. H. (2019). AffectNet: A database for facial expression, valence, and arousal computing in the wild. \u003cem\u003eIEEE Transactions on Affective Computing\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1), 18\u0026ndash;31. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/TAFFC.2017.2740923\u003c/span\u003e\u003cspan address=\"10.1109/TAFFC.2017.2740923\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNiedenthal, P. M., Mermillod, M., Maringer, M., \u0026amp; Hess, U. (2010). The simulation of smiles (SIMS) model: Embodied simulation and the meaning of facial expression. \u003cem\u003eBehavioral and Brain Sciences\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(6), 417\u0026ndash;433. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1017/S0140525X10000865\u003c/span\u003e\u003cspan address=\"10.1017/S0140525X10000865\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlszanowski, M., Pochwatko, G., Kuklinski, K., Scibor-Rylski, M., Lewinski, P., \u0026amp; Ohme, R. K. (2015). Warsaw set of emotional facial expression pictures: A validation study of facial display photographs. \u003cem\u003eFrontiers in Psychology\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e, 1516. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fpsyg.2014.01516\u003c/span\u003e\u003cspan address=\"10.3389/fpsyg.2014.01516\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlant, E. A., Hyde, J. S., Keltner, D., \u0026amp; Devine, P. G. (2000). The gender stereotyping of emotions. \u003cem\u003ePsychology of Women Quarterly\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e(1), 81\u0026ndash;92. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1471-6402.2000.tb01024.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1471-6402.2000.tb01024.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePochedly, J. T., Widen, S. C., \u0026amp; Russell, J. A. (2012). What emotion does the facial expression of disgust express? \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(6), 1315\u0026ndash;1319. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0027998\u003c/span\u003e\u003cspan address=\"10.1037/a0027998\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePowell, W. R., \u0026amp; Schirillo, J. A. (2009). Asymmetrical facial expressions in portraits and hemispheric laterality: A literature review. \u003cem\u003eLaterality\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(6), 545\u0026ndash;572. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/13576500802680336\u003c/span\u003e\u003cspan address=\"10.1080/13576500802680336\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePumarola, A., Agudo, A., Martinez, A. M., Sanfeliu, A., \u0026amp; Moreno-Noguer, F. (2020). GANimation: One-shot anatomically consistent facial animation. \u003cem\u003eInternational Journal of Computer Vision\u003c/em\u003e, \u003cem\u003e128\u003c/em\u003e(3), 698\u0026ndash;713. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11263-019-01210-3\u003c/span\u003e\u003cspan address=\"10.1007/s11263-019-01210-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRai, A., Gupta, H., Pandey, A., Carrasco, F. V., Takagi, S. J., Aubel, A., Kim, D., Prakash, A., \u0026amp; De La Torre, F. (2024). Towards realistic generative 3D face models. In \u003cem\u003e2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)\u003c/em\u003e (pp. 3726\u0026ndash;3736). IEEE. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/WACV57701.2024.00370\u003c/span\u003e\u003cspan address=\"10.1109/WACV57701.2024.00370\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRellecke, J., Sommer, W., \u0026amp; Schacht, A. (2012). Does processing of emotional facial expressions depend on intention? Time-resolved evidence from event-related brain potentials. \u003cem\u003eBiological Psychology\u003c/em\u003e, \u003cem\u003e90\u003c/em\u003e(1), 23\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.biopsycho.2012.02.002\u003c/span\u003e\u003cspan address=\"10.1016/j.biopsycho.2012.02.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoesch, E., Tamarit, L., Reveret, L., Grandjean, D., Sander, D., \u0026amp; Scherer, K. R. (2011). FACSGen: A tool to synthesize emotional facial expressions through systematic manipulation of facial action units. \u003cem\u003eJournal of Nonverbal Behavior\u003c/em\u003e, \u003cem\u003e35\u003c/em\u003e(1), 1\u0026ndash;16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10919-010-0095-9\u003c/span\u003e\u003cspan address=\"10.1007/s10919-010-0095-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRozin, P., Lowery, L., \u0026amp; Ebert, R. (1994). Varieties of disgust faces and the structure of disgust. \u003cem\u003eJournal of Personality and Social Psychology\u003c/em\u003e, \u003cem\u003e66\u003c/em\u003e(5), 870\u0026ndash;881. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0022-3514.66.5.870\u003c/span\u003e\u003cspan address=\"10.1037/0022-3514.66.5.870\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR Core Team. (2024). \u003cem\u003eR: A language and environment for statistical computing\u003c/em\u003e. R Foundation for Statistical Computing.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRychlowska, M., Jack, R. E., Garrod, O. G. B., Schyns, P. G., Martin, J. D., \u0026amp; Niedenthal, P. M. (2017). Functional smiles: Tools for love, sympathy, and war. \u003cem\u003ePsychological Science\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(9), 1259\u0026ndash;1270. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0956797617706082\u003c/span\u003e\u003cspan address=\"10.1177/0956797617706082\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSauter, D. A. (2010). More than happy: The need for disentangling positive emotions. \u003cem\u003eCurrent Directions in Psychological Science\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(1), 36\u0026ndash;40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0963721409359290\u003c/span\u003e\u003cspan address=\"10.1177/0963721409359290\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidtmann, G., Jennings, B. J., Sandra, D. A., Pollock, J., \u0026amp; Gold, I. (2020). The McGill Face Database: Validation and insights into the recognition of facial expressions of complex mental states. \u003cem\u003ePerception\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e(3), 310\u0026ndash;329. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0301006620901671\u003c/span\u003e\u003cspan address=\"10.1177/0301006620901671\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShrout, P. E., \u0026amp; Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. \u003cem\u003ePsychological Bulletin\u003c/em\u003e, \u003cem\u003e86\u003c/em\u003e(2), 420\u0026ndash;428. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0033-2909.86.2.420\u003c/span\u003e\u003cspan address=\"10.1037/0033-2909.86.2.420\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStrohminger, N., Gray, K., Chituc, V., Heffner, J., Schein, C., \u0026amp; Heagins, T. B. (2016). The MR2: A multi-racial, mega-resolution database of facial stimuli. \u003cem\u003eBehavior Research Methods\u003c/em\u003e, \u003cem\u003e48\u003c/em\u003e(3), 1197\u0026ndash;1204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/s13428-015-0641-9\u003c/span\u003e\u003cspan address=\"10.3758/s13428-015-0641-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., \u0026amp; Ortega-Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. \u003cem\u003eInformation Fusion\u003c/em\u003e, \u003cem\u003e64\u003c/em\u003e, 131\u0026ndash;148. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.inffus.2020.06.014\u003c/span\u003e\u003cspan address=\"10.1016/j.inffus.2020.06.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTottenham, N., Tanaka, J. W., Leon, A. C., McCarry, T., Nurse, M., Hare, T. A., Marcus, D. J., Westerlund, A., Casey, B., \u0026amp; Nelson, C. (2009). The NimStim set of facial expressions: Judgments from untrained research participants. \u003cem\u003ePsychiatry Research\u003c/em\u003e, \u003cem\u003e168\u003c/em\u003e(3), 242\u0026ndash;249. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.psychres.2008.05.006\u003c/span\u003e\u003cspan address=\"10.1016/j.psychres.2008.05.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTracy, J. L., Robins, R. W., \u0026amp; Schriber, R. A. (2009). Development of a FACS-verified set of basic and self-conscious emotion expressions. \u003cem\u003eEmotion\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(4), 554\u0026ndash;559. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/a0015766\u003c/span\u003e\u003cspan address=\"10.1037/a0015766\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWagner, H. L. (1993). On measuring performance in category judgment studies of nonverbal behavior. \u003cem\u003eJournal of Nonverbal Behavior\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(1), 3\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF00987006\u003c/span\u003e\u003cspan address=\"10.1007/BF00987006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, Z., Yaoma, K., Wu, Y., Burns, E., Sun, M., \u0026amp; Ying, H. (2024). Other ethnicity effects in ensemble coding of facial expressions. \u003cem\u003eAttention Perception \u0026amp; Psychophysics\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3758/s13414-024-02920-8\u003c/span\u003e\u003cspan address=\"10.3758/s13414-024-02920-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e Please see \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://osf.io/xh298\u003c/span\u003e\u003cspan address=\"https://osf.io/xh298\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e for more details.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePlease see \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/JinGao1997/Facial-Expressions-Rating-Task\u003c/span\u003e\u003cspan address=\"https://github.com/JinGao1997/Facial-Expressions-Rating-Task\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e for more details\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"psychological-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"prpf","sideBox":"Learn more about [Psychological Research](http://link.springer.com/journal/426)","snPcode":"426","submissionUrl":"https://submission.nature.com/new-submission/426/3","title":"Psychological Research","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"facial expressions, emotional stimuli, emotion recognition, FACSGen 2.0, social smiles","lastPublishedDoi":"10.21203/rs.3.rs-6908842/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6908842/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDespite their crucial role in emotion research, facial expression databases primarily contain stimuli of basic emotions, rather than subtle, socially nuanced ones. Here, we introduce a rigorous yet easily applicable pipeline for synthesizing fine-grained emotional expressions\u0026mdash;specifically, reward, affiliative, and dominance smiles\u0026mdash;from neutral photographs, guided by Facial Action Coding System (FACS) principles and the perspective that expressions serve as social signals. From neutral images of 90 individuals (45 female, 45 male), we generated five expressions each (including neutral and disgust), and mirrored each image to address hemiface biases. In an online validation study, 13 participants rated each image on emotional content, arousal, and plausibility, yielding 26 ratings per expression and model. Rating results show that (1) reward smiles were most reliably recognized, while affiliative and dominance smiles tended to be confused with related expressions and that (2) arousal and plausibility ratings varied systematically across expression types and model gender. In summary, the suggested expression generation pipeline and the validated stimuli offer a robust method and a scalable dataset for research on fine-grained emotion recognition and emotional communication.\u003c/p\u003e","manuscriptTitle":"Shades of Smiles: Creating Variants of Smiles from Neutral Images of Real Individuals - Method and Validation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-02 12:46:47","doi":"10.21203/rs.3.rs-6908842/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-21T23:47:10+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-15T18:46:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"212398623195501067362292636641109110706","date":"2025-07-24T16:07:55+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-23T07:51:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"30456083998488457015730620394017170097","date":"2025-07-16T00:52:52+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-23T11:05:01+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-18T06:03:46+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-17T04:42:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"Psychological Research","date":"2025-06-16T22:58:25+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"psychological-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"prpf","sideBox":"Learn more about [Psychological Research](http://link.springer.com/journal/426)","snPcode":"426","submissionUrl":"https://submission.nature.com/new-submission/426/3","title":"Psychological Research","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"fbdd19a7-4166-47be-8987-38b34e6a3842","owner":[],"postedDate":"July 2nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-23T16:01:05+00:00","versionOfRecord":{"articleIdentity":"rs-6908842","link":"https://doi.org/10.1007/s00426-026-02263-z","journal":{"identity":"psychological-research","isVorOnly":false,"title":"Psychological Research"},"publishedOn":"2026-03-19 15:58:29","publishedOnDateReadable":"March 19th, 2026"},"versionCreatedAt":"2025-07-02 12:46:47","video":"","vorDoi":"10.1007/s00426-026-02263-z","vorDoiUrl":"https://doi.org/10.1007/s00426-026-02263-z","workflowStages":[]},"version":"v1","identity":"rs-6908842","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6908842","identity":"rs-6908842","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}