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
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 |
|---|---|---|---|
| condition diffusion on masked input (DiffMAE) instead of plain DDPM pre-training | ViT-L · ImageNet-1k fine-tuning | top-1 accuracy (%): 83.4 -> 85.8 | Improved |
| plain diffusion pre-training (ADM) vs from scratch | ADM U-Net encoder · ImageNet-1k fine-tuning | top-1 accuracy (%): 82.0 -> 83.3 | No clear effect |
| use cross decoder instead of joint or cross-self decoder for recognition pre-training (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%) | Improved |
| use cross-self decoder instead of joint decoder for inpainting (ablation) | DiffMAE · IN-1K inpainting | LPIPS: 0.224 -> 0.216 | Improved |
| remove light-noise timesteps (use only t from 500 to 1000) (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): 84.7 -> 85.1 | Improved |
| increase noise via smaller rho (0.6) in schedule (ablation) | DiffMAE · IN-1K inpainting | LPIPS: 0.216 -> 0.228 | Hurt |
| predict pixels instead of noise (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): unstable -> 84.3 | Improved |
| add per-patch normalization to pixel targets (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): 84.3 -> 85.1 | Improved |
| mask patches (75%) instead of no masking (100%) (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): 83.4 -> 85.1 | Improved |
| random 75% mask instead of center mask for pre-training (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): 84.3 -> 85.1 | Improved |
| train with center mask instead of random mask for center-mask inpainting (ablation) | DiffMAE · IN-1K inpainting (center mask) | LPIPS: 0.142 -> 0.125 | Improved |
| vary decoder depth (4 to 12) or width (256 to 1024) (ablation) | DiffMAE · ImageNet-1k fine-tuning | top-1 accuracy (%): 85.0 / 85.1 / 85.1 (depth); 85.0 / 85.1 / 84.9 (width) | No clear effect |
Closest papers by meaning
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.91)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.91)
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.91)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.90)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
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.
