Research map / Video / Video generation
UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers
Min Zhao, Hongzhou Zhu, Yingze Wang, Bokai Yan, Jintao Zhang, Guande He, Ling Yang, Chongxuan Li, Jun Zhu
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 |
|---|---|---|---|
| suppress attention to out-of-window tokens (UltraViCo) vs best baseline RIFLEx | HunyuanVideo 4x · VBench 4x extrapolation | Imaging Quality: 50.01 -> 66.54 | Improved |
| UltraViCo vs prior methods on Wan beyond 3x | Wan2.1-1.3B · VBench 4x extrapolation | Dynamic Degree: 11 -> 47 | Improved |
| UltraViCo vs PE on content repetition | HunyuanVideo 4x · VBench 4x extrapolation | NoRepeat Score: 31.41 -> 99.87 | Improved |
| UltraViCo on overall consistency vs best baseline at 4x | HunyuanVideo 4x · VBench 4x extrapolation | Consistency: 0.9915 -> 0.9468 | Hurt |
| mask tokens at harmonic alignment positions (ablation) | HunyuanVideo · periodic attention | repetition and quality | Improved |
| increase masking ratio of out-of-window attention (ablation) | video DiT · controlled experiment | video quality | Improved |
| use parabolic or linear decay instead of constant decay (ablation) | UltraViCo · Wan 3x extrapolation | quality | No clear effect |
| lower decay factor alpha below 0.9 (ablation) | UltraViCo · Wan 3x extrapolation | consistency | Hurt |
| lower beta below 0.6 at harmonic positions (ablation) | UltraViCo · HunyuanVideo | consistency | Hurt |
| add UltraViCo to sliding window (ablation) | Sliding Window · Wan 6x extrapolation | Consistency: 0.8478 -> 0.9183 | Improved |
| add UltraViCo to FreeNoise (ablation) | FreeNoise · Wan 6x extrapolation | Consistency: 0.9243 -> 0.9431 | Improved |
| add UltraViCo to FIFO-Diffusion (ablation) | FIFO-Diffusion · Wan 6x extrapolation | Consistency: 0.9131 -> 0.9319 | Improved |
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.
- RIFLEx — RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers (ICML 2025) · better on 10 of 10 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
- RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers (ICML 2025 · similarity 0.91)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers (NeurIPS 2026 · similarity 0.89)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.89)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · 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.
