Research map / Image generation / Fast diffusion sampling
Perception Prioritized Training of Diffusion Models
Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, Sungroh Yoon
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 |
|---|---|---|---|
| Replace baseline loss weighting with P2 weighting | ADM diffusion models · FFHQ, CUB, AFHQ-Dog, Oxford Flowers, and MetFaces | FID-50k: 7.86 -> 6.92; 9.60 -> 6.95; 12.47 -> 11.55; 20.01 -> 17.29; 44.34 -> 36.80 | Improved |
| Replace baseline loss weighting with P2 weighting | ADM diffusion models · FFHQ, CUB, AFHQ-Dog, Oxford Flowers, and MetFaces | KID-50k (×103): 3.85 -> 3.46; 3.49 -> 2.38; 4.79 -> 4.10; 16.8 -> 14.8; 22.1 -> 17.6 | Improved |
| Use P2 weighting instead of baseline weighting | ADM diffusion model · CelebA-HQ 256×256 | FID: 7.22 -> 6.91 | Improved |
| Use P2 weighting instead of baseline weighting | ADM diffusion model · Oxford Flowers 256×256 | FID: 19.60 -> 17.29 | Improved |
| Use P2 weighting instead of baseline weighting | ADM diffusion model · FFHQ 256×256 | FID: 8.41 -> 6.97; StyleGAN2: 3.73 | Mixed |
| Remove self-attention outside the 8×8 bottleneck (ablation) | ADM with P2 weighting · MetFaces, 800k images | FID-10k: 39.51 -> 43.09 | Hurt |
| Replace the BigGAN residual block with the residual block of Ho et al. (ablation) | ADM with P2 weighting · MetFaces, 800k images | FID-10k: 41.93 -> 47.37 | Hurt |
| Use two residual blocks (ablation) | ADM with P2 weighting · MetFaces, 800k images | FID-10k: 41.93 -> 42.06 | Hurt |
| Set learning rate to 2.5e−5 (ablation) | ADM with P2 weighting · MetFaces, 800k images | FID-10k: 41.93 -> 39.51 | Improved |
| Use more sampling steps near t = 0 (130-60-60) rather than 250 uniform steps (ablation) | Baseline ADM · FFHQ | FID-10k: 10.88 -> 10.62 | Improved |
| Use 60-130-60 rather than 250 uniform sampling steps (ablation) | Baseline ADM · FFHQ | FID-10k: 10.88 -> 12.23 | Hurt |
| Increase the P2 exponent γ above 2 (ablation) | P2 weighting · Image generation | sample quality | Hurt |
Closest papers by meaning
- Scaling Properties of Diffusion Models For Perceptual Tasks (CVPR 2025 · similarity 0.89)
- What Matters When Repurposing Diffusion Models for General Dense Perception Tasks? (ICLR 2025 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.88)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.88)
- Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception (ICLR 2025 · similarity 0.88)
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.
