Hugging Face Daily Papers · May 29, 2026 · 6 min read

Colored Noise Diffusion Sampling

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

Standard diffusion SDEs waste energy by injecting uniform white noise, ignoring the fact that models resolve low-frequency structures before high-frequency details. We introduce Colored Noise Sampling (CNS), plug-and-play solver that dynamically targets injected energy toward unresolved frequency bands instead.\n","updatedAt":"2026-05-29T05:39:40.142Z","author":{"_id":"647858ba256b62e2198f217e","avatarUrl":"/avatars/b07432f31d9c29785ac46a3cc0375fc5.svg","fullname":"Noam Issachar","name":"NoamIssachar","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8994204998016357},"editors":["NoamIssachar"],"editorAvatarUrls":["/avatars/b07432f31d9c29785ac46a3cc0375fc5.svg"],"reactions":[{"reaction":"👍","users":["ollieollie","HadarD"],"count":2}],"isReport":false}},{"id":"6a1a40860499e06634bc1a33","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":359,"isUserFollowing":false},"createdAt":"2026-05-30T01:42:30.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [Bridging Restoration and Generation Manifolds in One-Step Diffusion for Real-World Super-Resolution](https://huggingface.co/papers/2604.24136) (2026)\n* [Probability-Conserving Flow Guidance](https://huggingface.co/papers/2605.20079) (2026)\n* [Disentangling Generation and Regression in Stochastic Interpolants for Controllable Image Restoration](https://huggingface.co/papers/2605.21381) (2026)\n* [Frequency-Aware Flow Matching for High-Quality Image Generation](https://huggingface.co/papers/2604.15521) (2026)\n* [Multi-Scale Generative Modeling with Heat Dissipation Flow Matching](https://huggingface.co/papers/2605.19371) (2026)\n* [DiLO: Decoupling Generative Priors and Neural Operators via Diffusion Latent Optimization for Inverse Problems](https://huggingface.co/papers/2604.11375) (2026)\n* [Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation](https://huggingface.co/papers/2604.19141) (2026)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. \nThe following papers were recommended by the Semantic Scholar API \n<ul>\n<li><a href=\"https://huggingface.co/papers/2604.24136\">Bridging Restoration and Generation Manifolds in One-Step Diffusion for Real-World Super-Resolution</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.20079\">Probability-Conserving Flow Guidance</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.21381\">Disentangling Generation and Regression in Stochastic Interpolants for Controllable Image Restoration</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.15521\">Frequency-Aware Flow Matching for High-Quality Image Generation</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.19371\">Multi-Scale Generative Modeling with Heat Dissipation Flow Matching</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.11375\">DiLO: Decoupling Generative Priors and Neural Operators via Diffusion Latent Optimization for Inverse Problems</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.19141\">Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation</a> (2026)</li>\n</ul>\n Please give a thumbs up to this comment if you found it helpful!\n If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code><a href=\"/librarian-bot\">@librarian-bot</a> recommend</code>\n","updatedAt":"2026-05-30T01:42:30.232Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":359,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7117651104927063},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.30332","authors":[{"_id":"6a19249056b4bb14ec65d088","user":{"_id":"697b80f027319f7874015f8a","avatarUrl":"/avatars/1fe6cfea0be79a2901c334687512c01d.svg","isPro":false,"fullname":"Hadar Davidson","user":"HadarD","type":"user","name":"HadarD"},"name":"Hadar Davidson","status":"claimed_verified","statusLastChangedAt":"2026-05-29T08:49:44.348Z","hidden":false},{"_id":"6a19249056b4bb14ec65d089","name":"Noam Issachar","hidden":false},{"_id":"6a19249056b4bb14ec65d08a","name":"Sagie Benaim","hidden":false}],"publishedAt":"2026-05-28T00:00:00.000Z","submittedOnDailyAt":"2026-05-29T00:00:00.000Z","title":"Colored Noise Diffusion Sampling","submittedOnDailyBy":{"_id":"647858ba256b62e2198f217e","avatarUrl":"/avatars/b07432f31d9c29785ac46a3cc0375fc5.svg","isPro":false,"fullname":"Noam Issachar","user":"NoamIssachar","type":"user","name":"NoamIssachar"},"summary":"Diffusion models achieve state-of-the-art image synthesis, with their generative trajectories fundamentally exhibiting a spectral bias, resolving low-frequency global structures early and high-frequency fine details later. Conventional stochastic differential equation (SDE) solvers fail to account for this dynamic, naively injecting uniform white noise throughout the entire process and misusing the finite energy budget. In this work, we establish a mathematical framework that reconsiders SDE inference as a targeted, frequency-decoupled energy transfer. Leveraging this framework, we introduce Colored Noise Sampling (CNS), a novel, training-free stochastic solver. Rather than injecting uniform white noise, CNS utilizes a dynamic, timestep- and frequency-dependent schedule that more efficiently allocates injected energy toward structurally unresolved frequency bands. By actively exploiting the model's inherent spectral bias, CNS systematically steers the generated distribution toward the true data manifold. Extensive experiments demonstrate that CNS significantly outperforms standard ODE and SDE baselines as a strictly plug-and-play, inference-time sampler substitution across diverse architectures (SiT, JiT, FLUX). Compared to standard sampling on ImageNet-256, CNS achieves substantial unguided FID reductions, improving from 8.26 to 6.27 on SiT-XL/2, 32.39 to 26.69 on JiT-B/16, and 11.88 to 8.31 on JiT-H/16, while yielding consistent relative FID improvements with Classifier-Free Guidance. Project page is available at https://hadardavidson.github.io/CNS/.","upvotes":14,"discussionId":"6a19249056b4bb14ec65d08b","projectPage":"https://hadardavidson.github.io/CNS/","githubRepo":"https://github.com/hadardavidson/colored-noise-sampling","githubRepoAddedBy":"user","ai_summary":"Diffusion models exhibit spectral bias in image synthesis, and a new sampling method called Colored Noise Sampling addresses this by dynamically allocating energy based on frequency-dependent schedules, leading to improved image quality metrics.","ai_keywords":["diffusion models","spectral bias","stochastic differential equation","SDE solvers","white noise","frequency-dependent schedule","energy transfer","Colored Noise Sampling","ODE","FID","Classifier-Free Guidance"],"githubStars":22,"organization":{"_id":"65157bc51e7b9224c9c6d460","name":"HUJI-IL","fullname":"The Hebrew University of Jerusalem","avatar":"https://www.gravatar.com/avatar/fbf7c0844f4246fadde2c5ef9867ccaf?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"697b80f027319f7874015f8a","avatarUrl":"/avatars/1fe6cfea0be79a2901c334687512c01d.svg","isPro":false,"fullname":"Hadar Davidson","user":"HadarD","type":"user"},{"_id":"65744a2fe09de6aa74026d80","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65744a2fe09de6aa74026d80/kCxIKdeBJwAPKmvlm7fDP.jpeg","isPro":false,"fullname":"Itay Chachy","user":"ItayChachy","type":"user"},{"_id":"6345a9b9a8c2ff9f1377faab","avatarUrl":"/avatars/a5f2b999ef8b967b2af9f41afcd9d475.svg","isPro":false,"fullname":"Sagie Benaim","user":"sagiebenaim","type":"user"},{"_id":"67af1b9f1f23276864a35c3c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IT8-ZHy7DGdXO3R6OsJVt.png","isPro":false,"fullname":"Israel Ben David","user":"IsraelBenDavid","type":"user"},{"_id":"6735e0ae4fa652d9618eaf73","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/brcZLLgG2WHndKVbZXki5.png","isPro":false,"fullname":"David Shavin","user":"david-shavin","type":"user"},{"_id":"635ba0c637c6a2c12e2daef9","avatarUrl":"/avatars/9fc2932d9ace2715f540f896754ec7d2.svg","isPro":false,"fullname":"Ollie McCarthy","user":"ollieollie","type":"user"},{"_id":"64543a1ccd09ceba0e14ecfd","avatarUrl":"/avatars/d4f3aca9aa8bb4188f68ffd9e0d1f881.svg","isPro":false,"fullname":"Omer Benishu","user":"omerbenishu","type":"user"},{"_id":"689c5777740b3276357aa2f7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/3qnR_WGVyt4iynicV8sWH.png","isPro":false,"fullname":"Michael Joseph","user":"MJ9k","type":"user"},{"_id":"65c43b8e61c8e6d06ab4bd41","avatarUrl":"/avatars/c97b98252ec3a0e27ea4e561fc901042.svg","isPro":false,"fullname":"NivCohen","user":"NivC","type":"user"},{"_id":"6671e6facda5ebe22f42b517","avatarUrl":"/avatars/f93089686656802131ae22ede9a3aed5.svg","isPro":false,"fullname":"Imri Shuval","user":"Imri-sh","type":"user"},{"_id":"647858ba256b62e2198f217e","avatarUrl":"/avatars/b07432f31d9c29785ac46a3cc0375fc5.svg","isPro":false,"fullname":"Noam Issachar","user":"NoamIssachar","type":"user"},{"_id":"644d8340d185572dd1e728f5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/UGVKlBiDLrQ5Qj1fSbSxP.png","isPro":false,"fullname":"Jesse Katz","user":"JesseK1627","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"65157bc51e7b9224c9c6d460","name":"HUJI-IL","fullname":"The Hebrew University of Jerusalem","avatar":"https://www.gravatar.com/avatar/fbf7c0844f4246fadde2c5ef9867ccaf?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.30332.md"}">

Papers

arxiv:2605.30332

Colored Noise Diffusion Sampling

Published on May 28

· Submitted by

Noam Issachar on May 29

The Hebrew University of Jerusalem

Upvote

Authors:

Hadar Davidson ,

Abstract

Diffusion models exhibit spectral bias in image synthesis, and a new sampling method called Colored Noise Sampling addresses this by dynamically allocating energy based on frequency-dependent schedules, leading to improved image quality metrics.

AI-generated summary

Diffusion models achieve state-of-the-art image synthesis, with their generative trajectories fundamentally exhibiting a spectral bias, resolving low-frequency global structures early and high-frequency fine details later. Conventional stochastic differential equation (SDE) solvers fail to account for this dynamic, naively injecting uniform white noise throughout the entire process and misusing the finite energy budget. In this work, we establish a mathematical framework that reconsiders SDE inference as a targeted, frequency-decoupled energy transfer. Leveraging this framework, we introduce Colored Noise Sampling (CNS), a novel, training-free stochastic solver. Rather than injecting uniform white noise, CNS utilizes a dynamic, timestep- and frequency-dependent schedule that more efficiently allocates injected energy toward structurally unresolved frequency bands. By actively exploiting the model's inherent spectral bias, CNS systematically steers the generated distribution toward the true data manifold. Extensive experiments demonstrate that CNS significantly outperforms standard ODE and SDE baselines as a strictly plug-and-play, inference-time sampler substitution across diverse architectures (SiT, JiT, FLUX). Compared to standard sampling on ImageNet-256, CNS achieves substantial unguided FID reductions, improving from 8.26 to 6.27 on SiT-XL/2, 32.39 to 26.69 on JiT-B/16, and 11.88 to 8.31 on JiT-H/16, while yielding consistent relative FID improvements with Classifier-Free Guidance. Project page is available at https://hadardavidson.github.io/CNS/.