AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Can We Achieve Efficient Diffusion Without Self-Attention? Distilling Self-Attention into Convolutions

Ziyi Dong, Chengxing Zhou, Weijian Deng, Pengxu Wei, Xiangyang Ji, Liang Lin

ICCV 2025 · Architectures and efficiency › Attention and transformers · approach: Attention head and block redesign · 2 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
replace self-attention with ΔConvBlockSD1.5 · LAION 10k 512x512DINOv2 Score: 42.74 -> 44.72Improved
replace self-attention with ΔConvBlockSDXL · LAION 10k 1024x1024DINOv2 Score: 42.65 -> 45.14Improved
replace self-attention with ΔConvBlockPixArt 512x512 · LAION 10kDINOv2 Score: 41.31 -> 42.56Improved
replace self-attention with ΔConvBlockSD1.5 · LAION 10k 512x512FDD: 210.86 -> 200.15Hurt
replace self-attention with ΔConvBlockPixArt 512x512 · LAION 10kFDD: 173.88 -> 181.06Hurt
replace self-attention with ΔConvBlockPixArt 1024x1024 · LAION 10kCLIP Score: 30.60 -> 30.55No clear effect
replace self-attention with ΔConvBlockSDXL · LAION 10k 1024x1024CLIP Score: 30.92 -> 30.87No clear effect
replace self-attention with ΔConvBlock (K=13)PixArt · 1024x1024 latencylatency (ms): 5.91 -> 3.53Improved
use ΔConvBlock instead of LinFusionPixArt · 1024x1024 latencylatency (ms): 5.82 -> 3.53Improved
increase kernel size K from 13 to 25 (ablation)ΔConvFusion · 4K latencylatency (ms): 54.84 -> 131.71Hurt
replace self-attention with ΔConvBlockSD1.5 · 16K FLOPsMean FLOPs (G): 11,010,934.11 -> 1,589.10Improved
replace self-attention with ΔConvBlockPixArt · 512x512 FLOPsMean FLOPs (G): 31.42 -> 8.43Improved

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