Try our model here: <a href=\"https://huggingface.co/zhuhz22/Causal-Forcing\">https://huggingface.co/zhuhz22/Causal-Forcing</a><br>And the full-stack open-source code: <a href=\"https://github.com/thu-ml/Causal-Forcing\" rel=\"nofollow\">https://github.com/thu-ml/Causal-Forcing</a><br>We release <strong>2-step frame-wise</strong> AR model with <strong>50% latency and even better quality</strong> compared to 4-step chunk-wise models!</p>\n","updatedAt":"2026-05-15T10:11:35.234Z","author":{"_id":"64c269a52d73768f07ac266c","avatarUrl":"/avatars/d497a960f8aef6a974907b68ed750c1c.svg","fullname":"Zhu Hongzhou","name":"zhuhz22","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":11,"isUserFollowing":false}},"numEdits":2,"identifiedLanguage":{"language":"en","probability":0.6453540325164795},"editors":["zhuhz22"],"editorAvatarUrls":["/avatars/d497a960f8aef6a974907b68ed750c1c.svg"],"reactions":[{"reaction":"🔥","users":["zhuhz22"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.15141","authors":[{"_id":"6a0680f4b1a8cbabc9f09877","name":"Min Zhao","hidden":false},{"_id":"6a0680f4b1a8cbabc9f09878","name":"Hongzhou Zhu","hidden":false},{"_id":"6a0680f4b1a8cbabc9f09879","name":"Kaiwen Zheng","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987a","name":"Zihan Zhou","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987b","name":"Bokai Yan","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987c","name":"Xinyuan Li","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987d","name":"Xiao Yang","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987e","name":"Chongxuan Li","hidden":false},{"_id":"6a0680f4b1a8cbabc9f0987f","name":"Jun Zhu","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/64c269a52d73768f07ac266c/f0lRysgW2mfGfbrSqXUQt.png"],"publishedAt":"2026-05-14T00:00:00.000Z","submittedOnDailyAt":"2026-05-15T00:00:00.000Z","title":"Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation","submittedOnDailyBy":{"_id":"64c269a52d73768f07ac266c","avatarUrl":"/avatars/d497a960f8aef6a974907b68ed750c1c.svg","isPro":false,"fullname":"Zhu Hongzhou","user":"zhuhz22","type":"user","name":"zhuhz22"},"summary":"Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \\ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the \\textbf{frame-wise 2-step setting} by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50\\% and Stage 2 training cost by sim4times. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .","upvotes":75,"discussionId":"6a0680f5b1a8cbabc9f09880","projectPage":"https://github.com/thu-ml/Causal-Forcing","ai_summary":"A novel causal consistency distillation method enables efficient frame-wise video generation with reduced latency and improved quality compared to existing chunk-wise approaches.","ai_keywords":["autoregressive diffusion","causal consistency distillation","causal CD","frame-wise autoregression","few-step AR initialization","diffusion distillation","VBench","VisionReward","Genie3","world model generation"],"organization":{"_id":"640d3084536d9fe0f005cac3","name":"thu-ml","fullname":"Tsinghua Machine Learning Group","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678587085174-633131798ef21f47308ce49b.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64c269a52d73768f07ac266c","avatarUrl":"/avatars/d497a960f8aef6a974907b68ed750c1c.svg","isPro":false,"fullname":"Zhu Hongzhou","user":"zhuhz22","type":"user"},{"_id":"644e3e5f030210812f413073","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg","isPro":true,"fullname":"Zilong Chen","user":"heheyas","type":"user"},{"_id":"63709c12f0fe906bdc50225e","avatarUrl":"/avatars/1c287b4ff6dc06b2812986b55d758906.svg","isPro":false,"fullname":"zzhzzh","user":"zzh0000","type":"user"},{"_id":"6350884a0f376d3c482bda54","avatarUrl":"/avatars/824c49aa2fb2e85801c001e2843c0576.svg","isPro":false,"fullname":"Wei Yu","user":"yuweiao","type":"user"},{"_id":"66d347eebb76fb26eedb256e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66d347eebb76fb26eedb256e/iCPF7GkmZu--XCsWzoucl.jpeg","isPro":false,"fullname":"tianqi liu","user":"tqliu","type":"user"},{"_id":"664da5a094234f3c17df8d3b","avatarUrl":"/avatars/db8f6988ae7ec289a210a747c54470ad.svg","isPro":false,"fullname":"Songsong Yu","user":"Two-hot","type":"user"},{"_id":"66f753556e6db689a2918903","avatarUrl":"/avatars/ff735ae95b58f04fb690eef84dc63854.svg","isPro":false,"fullname":"Abner Albert","user":"AbnerAGI","type":"user"},{"_id":"638f308fc4444c6ca870b60a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/638f308fc4444c6ca870b60a/Q11NK-8-JbiilJ-vk2LAR.png","isPro":true,"fullname":"Linoy Tsaban","user":"linoyts","type":"user"},{"_id":"63468720dd6d90d82ccf3450","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63468720dd6d90d82ccf3450/tVBFlmZNz8FRMkOrDaDID.jpeg","isPro":false,"fullname":"YSH","user":"BestWishYsh","type":"user"},{"_id":"66015e8aa4d296af07de538e","avatarUrl":"/avatars/a1295c631cc2646282c545859975ce4c.svg","isPro":false,"fullname":"Owen","user":"Owen777","type":"user"},{"_id":"6411c801e872ae3fb1e2c96e","avatarUrl":"/avatars/f8898dc13d700e545eedbbfab1c18353.svg","isPro":true,"fullname":"Franklin","user":"Franklinzhang","type":"user"},{"_id":"697b04bf95c7d6cade79b835","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/697b04bf95c7d6cade79b835/5urnq1O8o2Q5vgFIL1HTV.jpeg","isPro":false,"fullname":"Runjia Qian","user":"rickchains","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":2,"organization":{"_id":"640d3084536d9fe0f005cac3","name":"thu-ml","fullname":"Tsinghua Machine Learning Group","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678587085174-633131798ef21f47308ce49b.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.15141.md"}">
Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
Abstract
A novel causal consistency distillation method enables efficient frame-wise video generation with reduced latency and improved quality compared to existing chunk-wise approaches.
AI-generated summary
Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the \textbf{frame-wise 2-step setting} by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50\% and Stage 2 training cost by sim4times. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .
Community
This comment has been hidden (marked as Resolved) 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/2605.15141 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.15141 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.15141 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.