Analog Diffusion Models

Analog Diffusion Models | 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 Physical Sciences - Article Analog Diffusion Models Jiaqi Chu, Heiner Kremer, Fabian Falck, Grace Brennan, Burcu Canakci, and 14 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8919479/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract As generative artificial intelligence (GenAI) drives computational demands to unprecedented scales, digital hardware is approaching fundamental limits. Analog and optical systems promise orders-of-magnitude efficiency gains, but translating these to application-level gains is challenging due to the mismatch between hardware primitives and algorithmic requirements. Here, we introduce Analog Diffusion Models (ADMs) which implement diffusion inference with an implicit integration scheme, formulating each diffusion step as a fixed-point problem amenable for acceleration by efficient analog hardware. At the same time, training remains identical to that of conventional diffusion models, allowing adoption of established scalable training approaches with no additional overhead. We validate ADMs on analog hardware using three-dimensional optics with 2,304 programmable weights. On hardware, we generate two-dimensional distributions and latent-space distributions for MNIST, FashionMNIST, and ExtendedMNIST, demonstrating the feasibility of executing multi-layer diffusion processes entirely on noisy, non-traditional hardware. The current prototype reaches fixed-point convergence in 10–15 µs per diffusion step, with projections to nanosecond-scale convergence with miniaturization. In simulation, across multiple datasets, backbone architectures, and model sizes ranging from 32 million to 13 billion parameters, ADMs match the sample quality of standard methods with up to 16× fewer diffusion steps. Most importantly, they could achieve efficiency gains of more than 100× at the application level without sacrificing generation quality, 100× from hardware acceleration, and an additional 1-2× from algorithmic improvement, highlighting the multiplicative benefit of hardware–algorithm co-design. Together, these results establish ADMs as a scalable and general, hardware-aligned framework for low-latency and energy-efficient generative modeling on analog computing platforms. Physical sciences/Optics and photonics Physical sciences/Mathematics and computing Physical sciences/Physics generative AI analog computing diffusion models flow matching optical computing neuromorphic computing fixed-point compute paradigm Full Text Additional Declarations There is NO Competing Interest. Supplementary Files Supplementary.pdf Supplementary Materials Cite Share Download PDF Status: Under Review 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-8919479","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Physical Sciences - Article","associatedPublications":[],"authors":[{"id":607717860,"identity":"1dc5928e-5f9c-46ca-9982-abb41cc92bff","order_by":0,"name":"Jiaqi Chu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIie2QsQrCMBRFXwi0S7W4Cf5EQBAHaX8loeCmi0vHQkEXce9fVAQnh1cKdtHd4qAuLi5OUjfb6CSYrg45QwKXd8i7AdBo/hAzLA8ODgczAJQRRbVC3orHwcKPYng1SiCvhEObf6JaZUYvhzPgmOU3gc/NwLUNG6Hw14rFjG6fA07YcRQn8+tQRFMKZL47qrr0ypVQxKWSAqacnUKgZKpSzMdbyfdScdmWqpVWaFWvJCI+NKRC4jqlQ61JnzNPRLuqC8ouLFF1adrZMi98Ryyy/epeYPVj5HIu/N9KRRvYV4LKealoNBqNRsULMEdcRD+xiNgAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0008-4744-334X","institution":"Microsoft Research","correspondingAuthor":true,"prefix":"","firstName":"Jiaqi","middleName":"","lastName":"Chu","suffix":""},{"id":607717861,"identity":"6fe4c7d9-e86c-4829-b60d-cd87f5453bc8","order_by":1,"name":"Heiner Kremer","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Heiner","middleName":"","lastName":"Kremer","suffix":""},{"id":607717862,"identity":"7f51e821-6c4b-4db5-a2d1-e04535437fb6","order_by":2,"name":"Fabian Falck","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Fabian","middleName":"","lastName":"Falck","suffix":""},{"id":607717863,"identity":"4f54d02b-0bfa-4adf-a3e7-35465e5b60f5","order_by":3,"name":"Grace Brennan","email":"","orcid":"https://orcid.org/0000-0001-7081-0509","institution":"Microsoft research","correspondingAuthor":false,"prefix":"","firstName":"Grace","middleName":"","lastName":"Brennan","suffix":""},{"id":607717864,"identity":"83508d03-12e8-475c-9d39-6d13c020f5d7","order_by":4,"name":"Burcu Canakci","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Burcu","middleName":"","lastName":"Canakci","suffix":""},{"id":607717865,"identity":"dc112898-3e58-4a76-88bc-3764fa1044c3","order_by":5,"name":"James Clegg","email":"","orcid":"https://orcid.org/0009-0002-3428-7161","institution":"Microsoft Research Ltd","correspondingAuthor":false,"prefix":"","firstName":"James","middleName":"","lastName":"Clegg","suffix":""},{"id":607717866,"identity":"2e4a9b46-5b1b-4d55-aaa8-272ea3dd3c31","order_by":6,"name":"Daniel Cletheroe","email":"","orcid":"https://orcid.org/0009-0003-6444-9149","institution":"Microsoft Research (United Kingdom)","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Cletheroe","suffix":""},{"id":607717867,"identity":"873c1c8a-c012-41bc-a368-019c3cb04a71","order_by":7,"name":"Doug Kelly","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Doug","middleName":"","lastName":"Kelly","suffix":""},{"id":607717868,"identity":"2b388b76-19ec-4ffe-b9a5-c0ad4ec89fde","order_by":8,"name":"Christos Gkantsidis","email":"","orcid":"https://orcid.org/0000-0002-6898-2368","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Christos","middleName":"","lastName":"Gkantsidis","suffix":""},{"id":607717869,"identity":"330bb20f-dae2-43a0-bfe5-e5b6408a5459","order_by":9,"name":"Michael Hansen","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"","lastName":"Hansen","suffix":""},{"id":607717870,"identity":"a8db5c48-013a-4a5f-aa87-9e44e63f607b","order_by":10,"name":"Paul Jeha","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Paul","middleName":"","lastName":"Jeha","suffix":""},{"id":607717871,"identity":"df77bd1c-6622-4b9c-bc2d-3c4d2fcf40a7","order_by":11,"name":"Kirill Kalinin","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Kirill","middleName":"","lastName":"Kalinin","suffix":""},{"id":607717872,"identity":"6c977ee5-5043-4dfd-a973-1b0bb063b109","order_by":12,"name":"Jim Kleewein","email":"","orcid":"","institution":"Microsoft","correspondingAuthor":false,"prefix":"","firstName":"Jim","middleName":"","lastName":"Kleewein","suffix":""},{"id":607717873,"identity":"1e32bf77-423b-42d8-9aed-23b32ac1a2eb","order_by":13,"name":"Babak Rahmani","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Babak","middleName":"","lastName":"Rahmani","suffix":""},{"id":607717874,"identity":"d3bb6860-588f-4dc5-a9f5-9a2056130481","order_by":14,"name":"Saravan Rajmohan","email":"","orcid":"","institution":"Microsoft","correspondingAuthor":false,"prefix":"","firstName":"Saravan","middleName":"","lastName":"Rajmohan","suffix":""},{"id":607717875,"identity":"060fd1d9-2b44-42bd-939e-e09d7eb2d36f","order_by":15,"name":"Victor Rühle","email":"","orcid":"https://orcid.org/0000-0002-8957-7628","institution":"Microsoft","correspondingAuthor":false,"prefix":"","firstName":"Victor","middleName":"","lastName":"Rühle","suffix":""},{"id":607717876,"identity":"52a21c36-4e2a-424a-a5bd-9442ac9cce43","order_by":16,"name":"Jannes Gladrow","email":"","orcid":"","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Jannes","middleName":"","lastName":"Gladrow","suffix":""},{"id":607717877,"identity":"b2e18195-ebac-44a7-a5e8-12964b2f9de7","order_by":17,"name":"Francesca Parmigiani","email":"","orcid":"https://orcid.org/0000-0001-7784-2829","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Francesca","middleName":"","lastName":"Parmigiani","suffix":""},{"id":607717878,"identity":"5e3f3e68-ee0f-4a08-8f89-0ac17dbfc49b","order_by":18,"name":"Hitesh Ballani","email":"","orcid":"https://orcid.org/0000-0003-1573-3314","institution":"Microsoft Research","correspondingAuthor":false,"prefix":"","firstName":"Hitesh","middleName":"","lastName":"Ballani","suffix":""}],"badges":[],"createdAt":"2026-02-19 16:50:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8919479/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8919479/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105034331,"identity":"71a4c75c-0f24-4b9a-a01d-70a800b84b49","added_by":"auto","created_at":"2026-03-20 07:23:08","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7911831,"visible":true,"origin":"","legend":"Article File","description":"","filename":"Main.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8919479/v1_covered_871f4800-d50e-42b8-b81e-02efe077be49.pdf"},{"id":104873548,"identity":"5b068996-6700-4ea8-84a4-13b1d9ac05fd","added_by":"auto","created_at":"2026-03-18 08:27:10","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2995566,"visible":true,"origin":"","legend":"Supplementary Materials","description":"","filename":"Supplementary.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8919479/v1/9f62cc3caaef2dd501e32fb3.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Analog Diffusion Models","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"generative AI, analog computing, diffusion models, flow matching, optical computing, neuromorphic computing, fixed-point compute paradigm","lastPublishedDoi":"10.21203/rs.3.rs-8919479/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8919479/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"As generative artificial intelligence (GenAI) drives computational demands to unprecedented scales, digital hardware is approaching fundamental limits. Analog and optical systems promise orders-of-magnitude efficiency gains, but translating these to application-level gains is challenging due to the mismatch between hardware primitives and algorithmic requirements. Here, we introduce Analog Diffusion Models (ADMs) which implement diffusion inference with an implicit integration scheme, formulating each diffusion step as a fixed-point problem amenable for acceleration by efficient analog hardware. At the same time, training remains identical to that of conventional diffusion models, allowing adoption of established scalable training approaches with no additional overhead. We validate ADMs on analog hardware using three-dimensional optics with 2,304 programmable weights. On hardware, we generate two-dimensional distributions and latent-space distributions for MNIST, FashionMNIST, and ExtendedMNIST, demonstrating the feasibility of executing multi-layer diffusion processes entirely on noisy, non-traditional hardware. The current prototype reaches fixed-point convergence in 10–15 µs per diffusion step, with projections to nanosecond-scale convergence with miniaturization. In simulation, across multiple datasets, backbone architectures, and model sizes ranging from 32 million to 13 billion parameters, ADMs match the sample quality of standard methods with up to 16× fewer diffusion steps. Most importantly, they could achieve efficiency gains of more than 100× at the application level without sacrificing generation quality, 100× from hardware acceleration, and an additional 1-2× from algorithmic improvement, highlighting the multiplicative benefit of hardware–algorithm co-design. Together, these results establish ADMs as a scalable and general, hardware-aligned framework for low-latency and energy-efficient generative modeling on analog computing platforms.","manuscriptTitle":"Analog Diffusion Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-18 08:23:59","doi":"10.21203/rs.3.rs-8919479/v1","editorialEvents":[],"status":"published","journal":{"display":false,"email":"[email protected]","identity":"nature","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"nature","sideBox":"Learn more about [Nature](http://www.nature.com/nature/)","snPcode":"","submissionUrl":"","title":"Nature","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"70927461-ec45-4f01-84ba-df4e35f021bb","owner":[],"postedDate":"March 18th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"This content is not available.","date":"2026-05-02T14:36:47+00:00","index":3,"fulltext":"This content is not available."}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":64666170,"name":"Physical sciences/Optics and photonics"},{"id":64666171,"name":"Physical sciences/Mathematics and computing"},{"id":64666172,"name":"Physical sciences/Physics"}],"tags":[],"updatedAt":"2026-03-18T08:23:59+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-18 08:23:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8919479","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8919479","identity":"rs-8919479","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}