Research map / Video / Video generation
ReHyAt: Recurrent Hybrid Attention for Video Diffusion Transformers
Mohsen Ghafoorian, Amirhossein Habibian
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 |
|---|---|---|---|
| increase softmax chunk size Tc from 1 to 2 (ablation) | ReHyAt (25 blocks) · VBench 320x480 | Total: 80.97 -> 82.08 | Improved |
| increase chunk size Tc from 2 to 5 (ablation) | ReHyAt (25 blocks) · VBench 320x480 | Total: 82.08 -> 82.48 | Improved |
| enable chunk overlap To=1 versus 0 (ablation) | ReHyAt (25 blocks, Tc=3) · VBench 320x480 | Total: 81.56 -> 82.17 | Improved |
| increase overlap from To=1 to 3 (ablation) | ReHyAt · VBench 320x480 | Total: 82.17 -> 82.19 | No clear effect |
| make hybrid attention causal for RNN reformulation (ablation) | ReHyAt 15xTc=3,To=0 · VBench 320x480 | Total: 82.27 -> 82.35 | No clear effect |
| distill Wan2.1 into ReHyAt 15xTc=3,To=1 (ablation) | Wan2.1 1.3B · VBench | Semantic: 75.12 -> 80.70 | Improved |
| distill Wan2.1 into ReHyAt 15xTc=3,To=1 | Wan2.1 1.3B · VBench | Quality: 85.10 -> 84.57 | Hurt |
| ReHyAt versus original Wan2.1 in human preference | Wan2.1 1.3B · 500 paired comparisons | % preferred: 27.6 vs 29.0 (43.5 no preference) | No clear effect |
| ReHyAt versus Wan2.1 on VBench-2.0 | Wan2.1 1.3B · VBench-2.0 | Total: 56.0 -> 56.1 | No clear effect |
| replace flash attention with ReHyAt Tc=3 | Wan2.1 DiT blocks · Snapdragon8-Gen4, 121 frames | latency ms: 4809 -> 302 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SANA-Video — SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer (ICLR 2026) · better on 1 of 2 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
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.89)
- MobileWan: Closing the Quality Gap for Mobile Video Diffusion (NeurIPS 2026 · similarity 0.89)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.89)
- DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (CVPR 2025 · similarity 0.89)
- VORTA: Efficient Video Diffusion via Routing Sparse Attention (NeurIPS 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.
