Research map / Video / Video generation
FAST-AR: Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention
Dvir Samuel, Issar Tzachor, Matan Levy, Michael Green, Gal Chechik, Rami Ben-Ari
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 |
|---|---|---|---|
| use last key as group representative instead of mean key (ablation) | TempCache · Rolling-Forcing, attention recall | attention recall: 75% -> 90% | Improved |
| use last key as group representative instead of medoid key (ablation) | TempCache · Rolling-Forcing, attention recall | attention recall: 66% -> 90% | Improved |
| raise key-similarity merging threshold from 0.5 to 0.9 (ablation) | TempCache · Rolling-Forcing, attention recall | attention recall: 0.10 -> 0.90 | Improved |
| raise ANN quantization bit-width from 2-bit to 32-bit (ablation) | Quantized ANN matching · Rolling-Forcing, attention recall | attention recall: 0.20 -> 0.98 | Improved |
| raise ANN quantization bit-width from 2-bit to 32-bit (ablation) | Quantized ANN matching · Rolling-Forcing, throughput | FPS: 32 -> 10 | Hurt |
| replace FlowCache with TempCache-LSH for KV compression | Rolling-Forcing · LongVBench | component speedup: x2.3 -> x6.8 | Improved |
| replace RadialAttn with AnnSA-LSH for sparse self-attention | Rolling-Forcing · LongVBench | VBench: 61.51 -> 83.25 | Improved |
| combine TempCache, AnnSA and AnnCA instead of FlowCache+RadialAttn | Rolling-Forcing · LongVBench | E2E speedup: x4.4 -> x10.7 | Improved |
| combine TempCache, AnnSA and AnnCA instead of FlowCache+RadialAttn | LongVie2 · LongVGenBench | LongVGenBench score: 49.84 -> 64.91 | Improved |
| use ANN-based cross-attention pruning (AnnCA-Quant instead of LSH) | Rolling-Forcing · LongVBench | VBench: 83.23 -> 82.89 | No clear effect |
| use AnnSA-Quant instead of STA on short videos | HunyuanVideo (117 frames) · 5-second generation | latency (s): 719 -> 912 | Hurt |
| swap Dense FA3 for combined method on long rollouts | Rolling-Forcing · 3K-frame rollout | FPS and peak memory | Improved |
Closest papers by meaning
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.93)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.93)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.90)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.90)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.89)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 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.
