Research map / Video / Video generation
VMoBA: Mixture-of-Block Attention for Video Diffusion Models
Jianzong Wu, Liang Hou, Haotian Yang, Ye Tian, Pengfei Wan, Di ZHANG, Yunhai Tong
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 3D blocks to 1-2D partition (1-2-3D) (ablation) | VMoBA · Block partition ablation (Training 93x576x1024) | SubConsist: 86.12% -> 94.72% | Improved |
| add 2D blocks to 1-3D partition (ablation) | VMoBA · Block partition ablation (Training 93x576x1024) | SubConsist: 91.34% -> 94.72% | Improved |
| add 1D temporal blocks to 2-3D partition (ablation) | VMoBA · Block partition ablation (Training 93x576x1024) | SubConsist: 94.75% -> 94.72% | No clear effect |
| use global block selection instead of local (ablation) | VMoBA · Block selection ablation | SubConsist: 92.43% -> 94.72% | Improved |
| use threshold selection instead of top-k (global) (ablation) | VMoBA · Block selection ablation | SubConsist: 92.86% -> 94.72% | Improved |
| raise tau from 0.15 to 0.25 (ablation) | VMoBA · Threshold tau ablation | SubConsist: 91.12% -> 94.72% | Improved |
| raise tau from 0.25 to 0.50 (ablation) | VMoBA · Threshold tau ablation | SubConsist: 94.72% -> 94.42% | No clear effect |
| more blocks 24-48-144 vs 8-48-72 (ablation) | VMoBA · Number of blocks ablation | SubConsist: 94.72% -> 94.42% | No clear effect |
| fewer blocks 8-24-36 vs 8-48-72 (ablation) | VMoBA · Number of blocks ablation | SubConsist: 94.72% -> 93.59% | Hurt |
| VMoBA instead of MoBA in training | Wan 2.1 (93x576x1024) · Training | training time: 226 -> 187 GPU Hours | Improved |
| VMoBA instead of MoBA in training-free 81x480x832 | Wan 2.1 · Training-free 81x480x832 | latency: 126s -> 104s | Improved |
| VMoBA instead of full attention training | Wan 2.1 (93x576x1024) · Training | training time: 276 -> 187 GPU Hours | Improved |
Later papers that compared with this method
4 later accepted papers report VMoBA as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention (ICML 2026 · better on 4 of 4 shared results)
- MoGA: Mixture-of-Groups Attention for End-to-End Long Video Generation (ICLR 2026 · better on 4 of 4 shared results)
- VMonarch: Efficient Video Diffusion Transformers with Structured Attention (CVPR 2026 · better on 2 of 4 shared results)
- SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention (ICLR 2026 · better on 3 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 0.91)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.87)
- VORTA: Efficient Video Diffusion via Routing Sparse Attention (NeurIPS 2025 · similarity 0.87)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.87)
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.
