Research map / Video / Video generation
Stable Video Infinity: Infinite-Length Video Generation with Error Recycling
Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, Alexandre Alahi
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 |
|---|---|---|---|
| inject self-generated errors into the reference image (E_img) during fine-tuning (ablation) | SVI · Error term ablation (1-epoch tuning) | Image Quality: 57.63% -> 71.22% | Improved |
| add video latent error (E_vid) and noise error (E_noi) injection on top of E_img (ablation) | SVI · Error term ablation (1-epoch tuning) | Subject Consistency: 73.82% -> 94.69% | Improved |
| replace self-generated errors with handcrafted image degradations (ablation) | SVI · Self vs handcrafted errors | Image Quality: 56.97% -> 45.17% | Hurt |
| combine handcrafted errors with self-generated errors (ablation) | SVI · Self vs handcrafted errors | Image Quality: 56.97% -> 39.50% | Hurt |
| weaken error-recycling LoRA alpha from 1.0 to 0.2 at test time (ablation) | SVI · Error-recycling intensity (LoRA alpha) | Subject Consistency: 98.54% -> 92.16% | Hurt |
| increase error bank size Z beyond 500 (ablation) | SVI · Error bank size | Consistency and quality metrics | No clear effect |
| use a very small error bank (Z=1 or Z=10) (ablation) | SVI · Error bank size | Consistency and quality metrics | Hurt |
| apply SVI error-recycling fine-tuning to Wan 2.1 for ultra-long generation | Wan 2.1 · Ultra-long consistent video generation (250-sec) | Subject Consistency: 80.00% -> 97.50% | Improved |
| use SVI-Shot instead of FramePack | long video generation · Ultra-long creative video generation (250-sec) | Subject Consistency: 70.95% -> 91.96% | Improved |
| naive extension (copy reference image) used as a metric-fooling baseline | video evaluation · Ultra-long consistent video generation (250-sec) | Subject Consistency and Dynamic Degree: 100.00% consistency, 0.00% dynamic degree | Mixed |
Closest papers by meaning
- Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout (CVPR 2026 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.87)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.87)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.86)
- Generative Inbetweening through Frame-wise Conditions-Driven Video Generation (CVPR 2025 · similarity 0.86)
- InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation (NeurIPS 2025 · 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.
