Research map / Video / Video generation
Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention
Chengtao Lv, Yumeng Shi, Yushi Huang, Ruihao Gong, Shen Ren, Wenya Wang
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 finetuning to 1D sparse attention (ablation) | Self Forcing 1.3B, 90% sparsity · VBench | Total score: 73.0 -> 82.8 | Improved |
| add Chunk-Aware Growth to finetuned sparse attention (ablation) | Self Forcing 1.3B · VBench | Total score: 82.8 -> 83.2 | Mixed |
| add CAG: Dynamic Degree (ablation) | Self Forcing 1.3B · VBench | Dynamic degree: 46.4 -> 37.5 | Hurt |
| add Hierarchical Sparse Attention on top of CAG (ablation) | Self Forcing 1.3B · VBench | Total score: 83.2 -> 84.5 | Improved |
| add HSA: Dynamic Degree (ablation) | Self Forcing 1.3B · VBench | Dynamic degree: 37.5 -> 66.7 | Improved |
| vary HSA retrieved frames top-k 6, 9, 12 (ablation) | Light Forcing · VBench | Total score: 84.5, 84.4, 84.3 | No clear effect |
| apply moderate sparsity to first chunk only (ablation) | Self Forcing 1.3B · toy experiment (Fig. 2) | visual quality | Hurt |
| apply sparsity to chunks 2-7 with dense first chunk (ablation) | Self Forcing 1.3B · toy experiment (Fig. 2) | visual quality | No clear effect |
| replace dense attention with Light Forcing | Self Forcing 1.3B · VBench | Total score: 84.1 -> 84.5 | Improved |
| replace dense attention with Light Forcing | LongLive 1.3B · VBench | Total score: 83.2 -> 83.9 | Improved |
| replace dense attention with Light Forcing | Infinite-Forcing 1.3B · VBench-Long 15s | Total score: 83.6 -> 84.1 | Improved |
| add FP8 linear layers, LightVAE and fused kernels | Light Forcing · 5s video, RTX 5090 | end-to-end speedup: 3.08× | 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.
- VMoBA — VMoBA: Mixture-of-Block Attention for Video Diffusion Models (ICLR 2026) · better on 4 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
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.90)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.90)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.89)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · 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.
