By leveraging a causal history buffer to build consensus anchors, DiFA aligns the reverse inference trajectory directly with the forward diffusion process. This significantly reduces error accumulation during few-step sampling without any retraining costs. Demonstrating strong generality across various samplers and architectures, DiFA consistently enhances both sampling stability and fine texture details in pixel and latent spaces.</p>\n","updatedAt":"2026-07-21T02:46:55.288Z","author":{"_id":"647899155bf35e70ab5db9a0","avatarUrl":"/avatars/7e36645349436510486c15727c649a10.svg","fullname":"Shigui Li","name":"ShiguiLi","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.8467589020729065},"editors":["ShiguiLi"],"editorAvatarUrls":["/avatars/7e36645349436510486c15727c649a10.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.17972","authors":[{"_id":"6a5ed7cc4fe5d1d13e84aaeb","name":"Shigui Li","hidden":false},{"_id":"6a5ed7cc4fe5d1d13e84aaec","name":"Delu Zeng","hidden":false}],"publishedAt":"2026-07-20T00:00:00.000Z","submittedOnDailyAt":"2026-07-21T00:00:00.000Z","title":"DiFA: Inference-Time Forward-Process Alignment for Diffusion Models","submittedOnDailyBy":{"_id":"647899155bf35e70ab5db9a0","avatarUrl":"/avatars/7e36645349436510486c15727c649a10.svg","isPro":false,"fullname":"Shigui Li","user":"ShiguiLi","type":"user","name":"ShiguiLi"},"summary":"The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of the denoising process. In this work, we propose Forward-Process Aligned Diffusion prediction (DiFA), a training-free framework that reframes inference-time data prediction refinement as a sequential state estimation problem. Rather than reusing past outputs solely for numerical integration, DiFA treats iterative data predictions along the reverse trajectory as correlated observations to build a forward-aligned temporal consensus. Inspired by Kalman filtering, this consensus aggregates historical predictions according to structural consistency and noise-level compatibility. To counteract the over-smoothing tendency of temporal consensus, we introduce a deviation guidance mechanism to adaptively preserve residual details. Empirically, DiFA yields significant improvements on CIFAR-10 and ImageNet across the evaluated metrics, including FID, IS, and FD-DINOv2, demonstrating that aligning inference with the forward statistical structure substantially improves generative fidelity.","upvotes":1,"discussionId":"6a5ed7cc4fe5d1d13e84aaed","organization":{"_id":"65d9f8e26b8ab39009e13ba2","name":"cn-scut","fullname":"South China University of Technology","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/636608e78b6bd0d9f663f4cd/8oSX1nc7RpNDTBApOTzWL.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"65d9f8e26b8ab39009e13ba2","name":"cn-scut","fullname":"South China University of Technology","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/636608e78b6bd0d9f663f4cd/8oSX1nc7RpNDTBApOTzWL.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.17972.md","query":{}}">
DiFA: Inference-Time Forward-Process Alignment for Diffusion Models
Abstract
The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of the denoising process. In this work, we propose Forward-Process Aligned Diffusion prediction (DiFA), a training-free framework that reframes inference-time data prediction refinement as a sequential state estimation problem. Rather than reusing past outputs solely for numerical integration, DiFA treats iterative data predictions along the reverse trajectory as correlated observations to build a forward-aligned temporal consensus. Inspired by Kalman filtering, this consensus aggregates historical predictions according to structural consistency and noise-level compatibility. To counteract the over-smoothing tendency of temporal consensus, we introduce a deviation guidance mechanism to adaptively preserve residual details. Empirically, DiFA yields significant improvements on CIFAR-10 and ImageNet across the evaluated metrics, including FID, IS, and FD-DINOv2, demonstrating that aligning inference with the forward statistical structure substantially improves generative fidelity.
Community
By leveraging a causal history buffer to build consensus anchors, DiFA aligns the reverse inference trajectory directly with the forward diffusion process. This significantly reduces error accumulation during few-step sampling without any retraining costs. Demonstrating strong generality across various samplers and architectures, DiFA consistently enhances both sampling stability and fine texture details in pixel and latent spaces.
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/2607.17972 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.17972 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.17972 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.