Research map / Video / Video generation
End-to-End Training for Autoregressive Video Diffusion via Self-Resampling
Yuwei Guo, Ceyuan Yang, Hao He, Yang Zhao, Meng Wei, Zhenheng Yang, Weilin Huang, Dahua Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use autoregressive resampling instead of noise augmentation (ablation) | Resampling Forcing · VBench 15s (0-15s) | Temporal quality: 87.15 -> 90.46 | Improved |
| use autoregressive resampling instead of noise augmentation (ablation) | Resampling Forcing · VBench 15s (0-15s) | Visual quality: 61.90 -> 64.25 | Improved |
| use autoregressive resampling instead of noise augmentation (ablation) | Resampling Forcing · VBench 15s (0-15s) | Text alignment: 21.44 -> 25.26 | Improved |
| use autoregressive instead of parallel resampling (ablation) | Resampling Forcing · VBench 15s (0-15s) | Temporal quality: 88.01 -> 90.46 | Improved |
| use autoregressive instead of parallel resampling (ablation) | Resampling Forcing · VBench 15s (0-15s) | Visual quality: 62.51 -> 64.25 | Improved |
| use autoregressive instead of parallel resampling (ablation) | Resampling Forcing · VBench 15s (0-15s) | Text alignment: 24.51 -> 25.26 | Improved |
| use 75% sparse history routing instead of dense attention (ablation) | Resampling Forcing · VBench 15s (10-15s) | Temporal quality: 89.74 -> 87.03 | Hurt |
| use 75% sparse history routing instead of dense attention (ablation) | Resampling Forcing · VBench 15s (0-5s) | Temporal quality: 91.20 -> 90.18 | Hurt |
| reduce routing from top-5 to top-1 (95% sparsity) (ablation) | history routing · qualitative fish video | appearance consistency | Hurt |
| use dynamic history routing instead of sliding window of size 1 at equal sparsity (ablation) | history routing · qualitative fish video | appearance consistency | Improved |
| use very small timestep shifting factor s (ablation) | self-resampling · qualitative | error accumulation | Hurt |
| use very large timestep shifting factor s (ablation) | self-resampling · qualitative | semantic consistency | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.94)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.90)
- RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO (NeurIPS 2026 · similarity 0.90)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- History-Guided Video Diffusion (ICML 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
