<strong>MrFlow</strong> proposes a <strong>training-free</strong> multi-resolution strategy for accelerating image generation, following a clear coarse-to-fine pipeline: multi-step low-resolution structure sampling, pixel-space super-resolution, and one-step high-resolution detail refinement. This elegant design achieves faithful generation with up to <strong>10x end-to-end speedup</strong>, establishing a new SOTA among training-free diffusion acceleration methods. Moreover, MrFlow is orthogonal to pretrained timestep distillation methods, allowing straightforward combination and further pushing the end-to-end speedup beyond <strong>25x</strong>. Overall, the work is simple but effective.</p>\n","updatedAt":"2026-07-03T02:30:38.737Z","author":{"_id":"64612660933afb0106a9dee3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/Ea83e0zR_2m8foWy6J0AF.jpeg","fullname":"Xingyu Zheng","name":"Xingyu-Zheng","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.8358943462371826},"editors":["Xingyu-Zheng"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/Ea83e0zR_2m8foWy6J0AF.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.01642","authors":[{"_id":"6a4705276ee372f6920de24c","name":"Xingyu Zheng","hidden":false},{"_id":"6a4705276ee372f6920de24d","name":"Xianglong Liu","hidden":false},{"_id":"6a4705276ee372f6920de24e","name":"Yifu Ding","hidden":false},{"_id":"6a4705276ee372f6920de24f","name":"Weilun Feng","hidden":false},{"_id":"6a4705276ee372f6920de250","name":"Junqing Lin","hidden":false},{"_id":"6a4705276ee372f6920de251","name":"Jinyang Guo","hidden":false},{"_id":"6a4705276ee372f6920de252","name":"Haotong Qin","hidden":false}],"publishedAt":"2026-07-02T00:00:00.000Z","submittedOnDailyAt":"2026-07-03T00:00:00.000Z","title":"Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling","submittedOnDailyBy":{"_id":"64612660933afb0106a9dee3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/Ea83e0zR_2m8foWy6J0AF.jpeg","isPro":false,"fullname":"Xingyu Zheng","user":"Xingyu-Zheng","type":"user","name":"Xingyu-Zheng"},"summary":"Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, attaining more than 5x speedup without any training. However, the design of performing upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits the quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration while keeping OneIG within a 1% gap relative to that before acceleration, significantly surpassing other training-free acceleration strategies, and requiring no training or runtime dynamic identification whatsoever. MrFlow can further be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.","upvotes":15,"discussionId":"6a4705286ee372f6920de253","githubRepo":"https://github.com/Xingyu-Zheng/MrFlow","githubRepoAddedBy":"user","ai_summary":"MrFlow accelerates text-to-image diffusion by combining low-resolution generation with pixel-space super-resolution and noise injection, achieving up to 25x speedup without training or runtime modifications.","ai_keywords":["text-to-image diffusion","flow-matching models","multi-resolution generation","staged low-to-high-resolution pipeline","super-resolution","pretrained GAN-based model","noise injection","quadratic token reduction","timestep distillation"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":3},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64612660933afb0106a9dee3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/Ea83e0zR_2m8foWy6J0AF.jpeg","isPro":false,"fullname":"Xingyu Zheng","user":"Xingyu-Zheng","type":"user"},{"_id":"68d63e44ccf464a96ac18bcb","avatarUrl":"/avatars/01b93c353a7231fa9e053ba9edcb2b06.svg","isPro":false,"fullname":"Weilun Feng","user":"wlfeng","type":"user"},{"_id":"65c49589c0b1921e19260a8d","avatarUrl":"/avatars/7ce9af8c627f2a0c3db6bde82290ee1f.svg","isPro":false,"fullname":"Haotong Qin","user":"HaotongQin","type":"user"},{"_id":"66223ecd9b0c7e78df110684","avatarUrl":"/avatars/1c242806ca7f99ead78c01faa51f70a6.svg","isPro":false,"fullname":"Sherwin","user":"Sherwin-56","type":"user"},{"_id":"68a7e4621c8076361eafc1ee","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/rESOMEbVHYnd8hrrxvD70.png","isPro":false,"fullname":"lin","user":"Qlin00","type":"user"},{"_id":"65250f5d4f3337c6b1351298","avatarUrl":"/avatars/7ab2215e1e8866dd346259288be25b18.svg","isPro":false,"fullname":"HaoHaojie","user":"Levelower","type":"user"},{"_id":"6890abbe731f64ac24f8e1f3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/LJs8GBqES014gX-Em-OQL.png","isPro":false,"fullname":"Yifu Ding","user":"Yifu25","type":"user"},{"_id":"66fbf9bf63002e1de029ebbb","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66fbf9bf63002e1de029ebbb/dJj0Y7E-OLSryrWHsKZqC.jpeg","isPro":false,"fullname":"Zihao Jing","user":"zihaojing","type":"user"},{"_id":"64744cdfa51711a3b58b276e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64744cdfa51711a3b58b276e/I5rXqTA1phbNSZba5uosT.jpeg","isPro":false,"fullname":"YG","user":"AbSinthe07","type":"user"},{"_id":"62fa0d1b15e8b4d65a3471d4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1660554498046-noauth.jpeg","isPro":false,"fullname":"wenhao","user":"Wenhao-Sun","type":"user"},{"_id":"66618d062fc3b8663fad84e5","avatarUrl":"/avatars/115c379dd267a3cf107ae6db08f3ec3e.svg","isPro":false,"fullname":"kewei liao","user":"xuemolan","type":"user"},{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.01642.md","query":{}}">
Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling
Abstract
MrFlow accelerates text-to-image diffusion by combining low-resolution generation with pixel-space super-resolution and noise injection, achieving up to 25x speedup without training or runtime modifications.
Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, attaining more than 5x speedup without any training. However, the design of performing upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits the quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration while keeping OneIG within a 1% gap relative to that before acceleration, significantly surpassing other training-free acceleration strategies, and requiring no training or runtime dynamic identification whatsoever. MrFlow can further be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.
Community
MrFlow proposes a training-free multi-resolution strategy for accelerating image generation, following a clear coarse-to-fine pipeline: multi-step low-resolution structure sampling, pixel-space super-resolution, and one-step high-resolution detail refinement. This elegant design achieves faithful generation with up to 10x end-to-end speedup, establishing a new SOTA among training-free diffusion acceleration methods. Moreover, MrFlow is orthogonal to pretrained timestep distillation methods, allowing straightforward combination and further pushing the end-to-end speedup beyond 25x. Overall, the work is simple but effective.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2607.01642 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.01642 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.