Research map / Image generation / Fast diffusion sampling
Patch Diffusion: Faster and More Data-Efficient Training of Diffusion Models
Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou
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 |
|---|---|---|---|
| Train with patch-wise score matching and coordinate conditioning rather than full-image-only diffusion training | EDM-DDPM++ on CelebA-64×64 · CelebA-64×64 | FID: 1.66 -> 1.77 | Mixed |
| Train with patch-wise score matching and coordinate conditioning rather than full-image-only diffusion training | EDM-DDPM++ on FFHQ-64×64 · FFHQ-64×64 | FID: 2.60 -> 3.11 | Hurt |
| Train with patch-wise score matching and coordinate conditioning | EDM-DDPM++ on AFHQv2 small datasets · AFHQv2-Cat, AFHQv2-Dog, AFHQv2-Wild | FID: 4.60 -> 3.11; 4.94 -> 4.80; 2.59 -> 1.93 | Improved |
| Use patch-wise training | EDM-DDPM++ on CelebA-64×64 · CelebA-64×64 | Train-Time: ∼ 48h -> ∼ 24h | Improved |
| Use patch-wise training | LDM-ADM on LSUN-Bedroom-256×256 · LSUN-Bedroom-256×256 | FID: 4.32 -> 2.75 | Improved |
| Use patch-wise training | LDM-ADM on LSUN-Church-256×256 · LSUN-Church-256×256 | FID: 4.66 -> 2.66 | Improved |
| Use patch-wise training | LDM-ADM on LSUN-Bedroom-256×256 · LSUN-Bedroom-256×256 | Train-Time: ∼ 8 Days -> ∼ 4 days | Improved |
| Use patch-wise training | EDM-ADM on ImageNet-256×256 · ImageNet-256×256 | FID (with CFG): 10.56 -> 2.72 | Improved |
| Use stochastic rather than progressive patch-size scheduling (ablation) | Patch Diffusion · CelebA-64×64 | FID: 2.05 -> 1.66 | Improved |
| Use stochastic rather than progressive patch-size scheduling (ablation) | Patch Diffusion · FFHQ-64×64 | FID: 3.85 -> 3.11 | Improved |
| Include full-size images during patch training instead of using only patches (p=0.0) (ablation) | Patch Diffusion · CelebA-64×64 | FID | Improved |
| Increase the share of full-size images from p=0.0 to p=0.1 (ablation) | Patch Diffusion · CelebA-64×64 | FID | Improved |
Later papers that compared with this method
2 later accepted papers report Patch Diffusion as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- FaPS: A General and Fast Training Method for Diffusion Models (ICML 2026 · better on 6 of 6 shared results)
- WILD-Diffusion: A WDRO Inspired Training Method for Diffusion Models under Limited Data (ICLR 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diffusion Model Patching via Mixture-of-Prompts (AAAI 2025 · similarity 0.92)
- Patched Denoising Diffusion Models For High-Resolution Image Synthesis (ICLR 2024 · similarity 0.91)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.90)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.90)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.90)
- Plug-and-Play Diffusion Distillation (CVPR 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.
