Research map / Video / Video generation
Attention Surgery: An Efficient Recipe to Linearize Your Video Diffusion Transformer
Mohsen Ghafoorian, Denis Korzhenkov, 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 |
|---|---|---|---|
| add attention distillation to 15 linear-attention blocks (ablation) | Attention Surgery · VBench (ablation) | Total: 59.7 -> 78.9 | Improved |
| add attention distillation to 20xR8 hybrid (ablation) | Attention Surgery · VBench (ablation) | Total: 77.3 -> 80.0 | Improved |
| use heterogeneous block-rate selection instead of homogeneous (20xR8) (ablation) | Attention Surgery · VBench (ablation) | Total: 80.0 -> 80.9 | Improved |
| use heterogeneous block-rate selection instead of homogeneous (15xR4) (ablation) | Attention Surgery · VBench (ablation) | Total: 81.0 -> 81.9 | Improved |
| use heterogeneous block-rate selection instead of homogeneous (20xR4) (ablation) | Attention Surgery · VBench (ablation) | Total: 80.9 -> 81.1 | Improved |
| use heterogeneous block-rate selection instead of homogeneous (25xR4) (ablation) | Attention Surgery · VBench (ablation) | Total: 79.9 -> 80.2 | Improved |
| value distillation instead of attention distillation (20xR8) (ablation) | Attention Surgery · VBench (ablation) | Total: 79.5 -> 80.0 | Improved |
| value distillation instead of attention distillation (15xR8) (ablation) | Attention Surgery · VBench (ablation) | Total: 81.2 -> 80.1 | Hurt |
| increase phi complexity from poly degree 2 / 2 layers to degree 6 / 3 layers (ablation) | Attention Surgery · VBench (ablation) | Total (10xR8): 82.1 -> 82.1 | No clear effect |
| replace original Wan2.1 attention with 15xR2 hybrid attention surgery | Wan2.1 1.3B · VBench | Total: 83.10 -> 83.21 | No clear effect |
| replace original Wan2.1 with 15xR2 hybrid | Wan2.1 1.3B · VBench-2.0 | Total: 56.0 -> 55.1 | Hurt |
| replace original Wan2.1 with 15xR2 hybrid | Wan2.1 1.3B · Human preference study | preference: 31.0 vs 29.3 preferred; 39.7 no preference | No clear effect |
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 3 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
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.92)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.92)
- VORTA: Efficient Video Diffusion via Routing Sparse Attention (NeurIPS 2025 · similarity 0.91)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.90)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.90)
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.
