AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

Diffusion Models as Masked Autoencoders

Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, Haoqi Fan, Hu Xu, Huiyu Wang, Cihang Xie, Alan Yuille, Christoph Feichtenhofer

ICCV 2023 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 80 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
condition diffusion on masked input (DiffMAE) instead of plain DDPM pre-trainingViT-L · ImageNet-1k fine-tuningtop-1 accuracy (%): 83.4 -> 85.8Improved
plain diffusion pre-training (ADM) vs from scratchADM U-Net encoder · ImageNet-1k fine-tuningtop-1 accuracy (%): 82.0 -> 83.3No clear effect
use cross decoder instead of joint or cross-self decoder for recognition pre-training (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%)Improved
use cross-self decoder instead of joint decoder for inpainting (ablation)DiffMAE · IN-1K inpaintingLPIPS: 0.224 -> 0.216Improved
remove light-noise timesteps (use only t from 500 to 1000) (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): 84.7 -> 85.1Improved
increase noise via smaller rho (0.6) in schedule (ablation)DiffMAE · IN-1K inpaintingLPIPS: 0.216 -> 0.228Hurt
predict pixels instead of noise (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): unstable -> 84.3Improved
add per-patch normalization to pixel targets (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): 84.3 -> 85.1Improved
mask patches (75%) instead of no masking (100%) (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): 83.4 -> 85.1Improved
random 75% mask instead of center mask for pre-training (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): 84.3 -> 85.1Improved
train with center mask instead of random mask for center-mask inpainting (ablation)DiffMAE · IN-1K inpainting (center mask)LPIPS: 0.142 -> 0.125Improved
vary decoder depth (4 to 12) or width (256 to 1024) (ablation)DiffMAE · ImageNet-1k fine-tuningtop-1 accuracy (%): 85.0 / 85.1 / 85.1 (depth); 85.0 / 85.1 / 84.9 (width)No clear effect

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