AI Paper Atlas

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

ICML 2026 · Video › Video generation · approach: Efficient video diffusion models

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
use last key as group representative instead of mean key (ablation)TempCache · Rolling-Forcing, attention recallattention recall: 75% -> 90%Improved
use last key as group representative instead of medoid key (ablation)TempCache · Rolling-Forcing, attention recallattention recall: 66% -> 90%Improved
raise key-similarity merging threshold from 0.5 to 0.9 (ablation)TempCache · Rolling-Forcing, attention recallattention recall: 0.10 -> 0.90Improved
raise ANN quantization bit-width from 2-bit to 32-bit (ablation)Quantized ANN matching · Rolling-Forcing, attention recallattention recall: 0.20 -> 0.98Improved
raise ANN quantization bit-width from 2-bit to 32-bit (ablation)Quantized ANN matching · Rolling-Forcing, throughputFPS: 32 -> 10Hurt
replace FlowCache with TempCache-LSH for KV compressionRolling-Forcing · LongVBenchcomponent speedup: x2.3 -> x6.8Improved
replace RadialAttn with AnnSA-LSH for sparse self-attentionRolling-Forcing · LongVBenchVBench: 61.51 -> 83.25Improved
combine TempCache, AnnSA and AnnCA instead of FlowCache+RadialAttnRolling-Forcing · LongVBenchE2E speedup: x4.4 -> x10.7Improved
combine TempCache, AnnSA and AnnCA instead of FlowCache+RadialAttnLongVie2 · LongVGenBenchLongVGenBench score: 49.84 -> 64.91Improved
use ANN-based cross-attention pruning (AnnCA-Quant instead of LSH)Rolling-Forcing · LongVBenchVBench: 83.23 -> 82.89No clear effect
use AnnSA-Quant instead of STA on short videosHunyuanVideo (117 frames) · 5-second generationlatency (s): 719 -> 912Hurt
swap Dense FA3 for combined method on long rolloutsRolling-Forcing · 3K-frame rolloutFPS and peak memoryImproved

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