AI Paper Atlas

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

ICML 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add finetuning to 1D sparse attention (ablation)Self Forcing 1.3B, 90% sparsity · VBenchTotal score: 73.0 -> 82.8Improved
add Chunk-Aware Growth to finetuned sparse attention (ablation)Self Forcing 1.3B · VBenchTotal score: 82.8 -> 83.2Mixed
add CAG: Dynamic Degree (ablation)Self Forcing 1.3B · VBenchDynamic degree: 46.4 -> 37.5Hurt
add Hierarchical Sparse Attention on top of CAG (ablation)Self Forcing 1.3B · VBenchTotal score: 83.2 -> 84.5Improved
add HSA: Dynamic Degree (ablation)Self Forcing 1.3B · VBenchDynamic degree: 37.5 -> 66.7Improved
vary HSA retrieved frames top-k 6, 9, 12 (ablation)Light Forcing · VBenchTotal score: 84.5, 84.4, 84.3No clear effect
apply moderate sparsity to first chunk only (ablation)Self Forcing 1.3B · toy experiment (Fig. 2)visual qualityHurt
apply sparsity to chunks 2-7 with dense first chunk (ablation)Self Forcing 1.3B · toy experiment (Fig. 2)visual qualityNo clear effect
replace dense attention with Light ForcingSelf Forcing 1.3B · VBenchTotal score: 84.1 -> 84.5Improved
replace dense attention with Light ForcingLongLive 1.3B · VBenchTotal score: 83.2 -> 83.9Improved
replace dense attention with Light ForcingInfinite-Forcing 1.3B · VBench-Long 15sTotal score: 83.6 -> 84.1Improved
add FP8 linear layers, LightVAE and fused kernelsLight Forcing · 5s video, RTX 5090end-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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy