AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

DiTFastAttnV2: Head-wise Attention Compression for Multi-Modality Diffusion Transformers

Hanling Zhang, Rundong Su, Zhihang Yuan, Pengtao Chen, Mingzhu Shen, Yibo Fan, Shengen Yan, Guohao Dai, Yu Wang

ICCV 2025 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 14 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 output caching to arrow attention in the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4attention sparsity: 0.30 -> 0.55Improved
add output caching to arrow attention in the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4LPIPS: 0.275 -> 0.238Improved
add output caching to arrow attention in the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4HPSv2: 0.2943 -> 0.2935Hurt
add CFG sharing to the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4attention sparsity: 0.55 -> 0.54No clear effect
add residual sharing to the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4LPIPS: 0.238 -> 0.196Improved
add residual sharing to the method set (ablation)DiTFastAttnV2 (SD3) · MS-COCO 1K, delta=0.4HPSv2: 0.2935 -> 0.2906Hurt
replace sliding window with arrow attention in DiTFastAttnDiTFastAttn (SD3) · MS-COCO 1Kattention sparsity: 0.32 -> 0.44Improved
use head-wise DiTFastAttnV2 instead of layer-wise attention cachingAttention Caching (N=2) (SD3) · MS-COCO 1KLPIPS: 0.323 -> 0.238Improved
use head-wise DiTFastAttnV2 instead of layer-wise attention cachingAttention Caching (N=2) (SD3) · MS-COCO 1Kattention sparsity: 0.50 -> 0.55Improved
use single-layer RSE calibration with ILP plan search instead of full-output exhaustive searchDiTFastAttn compression plan search · FLUX 2K plan searchsearch time: over 10 hours -> 15 minutesImproved
use head-wise arrow attention kernel instead of FlexAttentionattention kernel · 16384 visual tokens, 25% sparsityspeedup vs FlashAttention2: 1.22x -> 1.44xImproved
use head-wise arrow attention kernel instead of FlexAttentionattention kernel · 16384 visual tokens, 75% sparsityspeedup vs FlashAttention2: 3.09x -> 3.55xImproved

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