Research map / Video / Video generation
VMonarch: Efficient Video Diffusion Transformers with Structured Attention
Cheng Liang, Haoxian Chen, Liang Hou, Qi Fan, Gangshan Wu, Xin Tao, Limin Wang
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 first-frame full-attention recomputation (ablation) | VMonarch · VBench training-free | PSNR vs full attention: 11.65 -> 12.59 | Improved |
| add first-frame recomputation (fine-tuned) (ablation) | VMonarch · VBench fine-tuned | validation loss: 1.27 -> 1.25 | Improved |
| replace recomputation with attention-sink averaging (ablation) | VMonarch · VBench fine-tuned | validation loss: 1.25 -> 1.78 | Hurt |
| increase Monarch iterations from 2 to 7 (fine-tuned) (ablation) | VMonarch · VBench fine-tuned | Dynamic Degree: 68.06 -> 36.11 | Hurt |
| increase Monarch iterations from 1 to 3 (training-free) (ablation) | VMonarch · VBench training-free | PSNR vs full attention: 11.18 -> 12.21 | Improved |
| use block size of two frames instead of one (ablation) | VMonarch · VBench fine-tuned | validation loss: 1.25 -> 1.37 | Hurt |
| fine-tune 1500 steps (training-free to tuned) (ablation) | VMonarch · VBench Dynamic Degree | Dynamic Degree: 54.17% -> 62.50% | Improved |
| VMonarch vs VSA at 90% sparsity, training-free | VMonarch · VBench training-free | Aesthetic Quality: 42.86% -> 63.65% | Improved |
| VMonarch vs VSA on end-to-end time (61x448x832) | VMonarch · Inference time | time (s): 49.9 -> 47.7 | Improved |
| VMonarch vs full attention on long video time (141 frames) | VMonarch · Inference time | time (s): 214.2 -> 102.7 | Improved |
| VMonarch vs full attention on Dynamic Degree after tuning | VMonarch · VBench 61x448x832 | Dynamic Degree: 68.06% -> 62.50% | Hurt |
| VMonarch vs VSA on Wan2.1-14B quality | VMonarch · VBench 93x704x1280 | Aesthetic Quality: 66.40% -> 65.91% | Mixed |
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.
- VMoBA — VMoBA: Mixture-of-Block Attention for Video Diffusion Models (ICLR 2026) · better on 2 of 4 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
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.92)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.91)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.91)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.90)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- 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.
