Research map / Video / Video generation
Pathwise Test-Time Correction for Autoregressive Long Video Generation
Xunzhi Xiang, Zixuan Duan, Guiyu Zhang, Haiyu Zhang, Zhe Gao, Junta Wu, Shaofeng Zhang, Tengfei Wang, Qi Fan, Chunchao Guo
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 |
|---|---|---|---|
| add pathwise test-time correction | Self-Forcing · VBench 30-second generation | Subject Consistency: 92.5 -> 94.0 | Improved |
| add pathwise test-time correction | CausVid · VBench 30-second generation | Subject Consistency: 91.2 -> 93.2 | Improved |
| add pathwise test-time correction | CausVid · VBench 30-second generation | Dynamic Degree: 50.8 -> 69.5 | Improved |
| add pathwise test-time correction | Self-Forcing · VBench 30-second generation | Dynamic Degree: 62.5 -> 60.2 | Hurt |
| add pathwise test-time correction | Self-Forcing · 30-second color shift | Color-shift L1: 1.028 -> 0.644 | Improved |
| add pathwise test-time correction | Self-Forcing · 30-second JEPA | JEPA Difference: 0.191 -> 0.170 | Improved |
| add pathwise test-time correction | CausVid · VBench 30-second generation | Motion Smoothness: 98.1 -> 97.6 | Hurt |
| use TTC instead of Best-of-N or Search-over-Path | Self-Forcing · 30-second test-time scaling | Subject Consistency: 92.4/92.7 -> 94.0 at 10.53 vs 3.16 fps | Improved |
| use pathwise instead of single-point correction (ablation) | TTC · Self-Forcing 30s | t-LPIPS boundary: 0.205 -> 0.176 | Improved |
| correct at timestep 750 only vs no correction (ablation) | TTC · Self-Forcing 30s | t-LPIPS boundary: 0.178 -> 0.161 | Improved |
| correct at timestep 500 or 250 only vs no correction (ablation) | TTC · Self-Forcing 30s | t-LPIPS boundary: 0.178 -> 0.182/0.183 | No clear effect |
| apply test-time optimization (LoRA) with reconstruction loss on early frames (ablation) | distilled AR model · toy experiment (qualitative) | drift and motion | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- Rolling Forcing — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (ICLR 2026) · better on 0 of 1 shared result
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 0 of 1 shared result
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
- Video-T1: Test-time Scaling for Video Generation (ICCV 2025 · similarity 0.89)
- RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO (NeurIPS 2026 · similarity 0.87)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.87)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.86)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.86)
- T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback (NeurIPS 2024 · similarity 0.86)
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.
