Qwen-Edit+: Scaling Image Editing with VLM-Guided Consistency and Aesthetic Preference Distillation

Qwen-Edit+: Scaling Image Editing with VLM-Guided Consistency and Aesthetic Preference Distillation | 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 Qwen-Edit+: Scaling Image Editing with VLM-Guided Consistency and Aesthetic Preference Distillation Fan Tang, Siyuan Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9352857/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Instruction-based image editing has advanced substantially with the emergence of Diffusion Transformers (DiTs). However, a central challenge remains unresolved: how to accurately execute complex editing instructions while preserving the structural consistency and visual quality of the source image. Existing methods are primarily limited by three factors: noisy and imbalanced training data, insufficient structural supervision, and inadequate alignment with human aesthetic preferences. To address these issues, we propose Qwen-Edit+, a unified framework for image editing. Specifically, we first introduce Semantic-Consistency Aware Filtering (SCAF) and Distribution-Adaptive Sampling (DAS) to construct high-quality and category-balanced training data. We then propose a VLM-aware Consistency Loss (VCL), which exploits the hierarchical hidden states of Qwen2.5-VL to provide deep semantic and structural supervision. Finally, we incorporate Aesthetic Preference Distillation (APD) to further improve visual harmony and perceptual quality. In comparative experiments, our method achieved a CLIP Score of 0.347, an LPIPS of 0.219, a PSNR of 25.63, and an Aesthetic Score of 6.31 on Qwen-Consistent-Edit-1.2K, outperforming representative baselines in editability, structural fidelity, and visual quality. Artificial Intelligence and Machine Learning instruction-based image editing vision-language models structural consistency aesthetic preference distillation data curation Full Text Additional Declarations The authors declare potential competing interests as follows: Competing Interests Both authors are affiliated with Chongqing Valiant Cat Technology Co., Ltd., which organized this research. The authors declare that they have no additional competing financial or non-financial interests. Cite Share Download PDF Status: Posted Version 1 posted 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-9352857","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":619422494,"identity":"ea1ec5f9-1ab1-4b5d-8c50-db20065e2ba8","order_by":0,"name":"Fan Tang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIiWNgGAWjYLCCCgYGOTb2xsYHQHYCQdU8IOIMA4MxP8/hZgOoFsYGYrQkzpyR3iZBlBZ79rOHXxyoucO44UBiWzVPhV0eg3Tz8Qd4beHJS7M4cOwZs8GBg223ec4kFzPIHEsk4LAcM+MPbIfZDA42tt3mbWNObJDIMcSvhf+NmcGBf4d5DA4zthXz/qsHasn/iF+LRI7xg4NthyUk2xjbmHkbDoNsIRBiN96YMRzsO2zAz8PYLDnn2PHENok0wxn4tLD35xh/OPDtcH2b/POHH97UVCf2SyQ/+IBPCxCwScBYTKBYYiOgHASY4WYy/iBC+SgYBaNgFIw8AACnClOBFvCTiwAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0008-7945-8693","institution":"AI Laboratory, Chongqing Valiant Cat Technology","correspondingAuthor":true,"prefix":"","firstName":"Fan","middleName":"","lastName":"Tang","suffix":""},{"id":619422495,"identity":"3c6d7c74-d491-470c-903d-d3e4301c7df6","order_by":1,"name":"Siyuan Li","email":"","orcid":"https://orcid.org/0000-0002-2354-4233","institution":"AI Laboratory, Chongqing Valiant Cat Technology","correspondingAuthor":false,"prefix":"","firstName":"Siyuan","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2026-04-08 07:08:02","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":true,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9352857/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9352857/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106724425,"identity":"577fdfb7-f98b-4adf-9d73-d0eb6d951609","added_by":"auto","created_at":"2026-04-12 18:28:00","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":463148,"visible":true,"origin":"","legend":"","description":"","filename":"MeituFinal.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9352857/v1_covered_b1d00dad-b0f3-4aec-85b0-e037a038f1de.pdf"}],"financialInterests":"The authors declare potential competing interests as follows: Competing Interests\nBoth authors are affiliated with Chongqing Valiant Cat Technology Co., Ltd., which organized this research. The authors declare that they have no additional competing financial or non-financial interests.","formattedTitle":"\u003cp\u003eQwen-Edit+: Scaling Image Editing with VLM-Guided Consistency and Aesthetic Preference Distillation\u003c/p\u003e","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Chongqing Valiant Cat Technology Co., Ltd.","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"instruction-based image editing, vision-language models, structural consistency, aesthetic preference distillation, data curation","lastPublishedDoi":"10.21203/rs.3.rs-9352857/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9352857/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eInstruction-based image editing has advanced substantially with the emergence of Diffusion Transformers (DiTs). However, a central challenge remains unresolved: how to accurately execute complex editing instructions while preserving the structural consistency and visual quality of the source image. Existing methods are primarily limited by three factors: noisy and imbalanced training data, insufficient structural supervision, and inadequate alignment with human aesthetic preferences. To address these issues, we propose Qwen-Edit+, a unified framework for image editing. Specifically, we first introduce Semantic-Consistency Aware Filtering (SCAF) and Distribution-Adaptive Sampling (DAS) to construct high-quality and category-balanced training data. We then propose a VLM-aware Consistency Loss (VCL), which exploits the hierarchical hidden states of Qwen2.5-VL to provide deep semantic and structural supervision. Finally, we incorporate Aesthetic Preference Distillation (APD) to further improve visual harmony and perceptual quality. In comparative experiments, our method achieved a CLIP Score of 0.347, an LPIPS of 0.219, a PSNR of 25.63, and an Aesthetic Score of 6.31 on Qwen-Consistent-Edit-1.2K, outperforming representative baselines in editability, structural fidelity, and visual quality.\u003c/p\u003e","manuscriptTitle":"Qwen-Edit+: Scaling Image Editing with VLM-Guided Consistency and Aesthetic Preference Distillation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-09 03:01:24","doi":"10.21203/rs.3.rs-9352857/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3f1e9e62-e7b0-40ed-88f7-a2a7d62cacc0","owner":[],"postedDate":"April 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65912161,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-04-09T03:01:24+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-09 03:01:24","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9352857","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9352857","identity":"rs-9352857","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}