Research map / Image generation / Fast diffusion sampling
Spectral Prior for Reducing Exposure Bias in Diffusion Models
Yuya Kobayashi, Masato Ishii, Yuhta Takida, Takashi Shibuya, Yuki Mitsufuji
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 |
|---|---|---|---|
| add SPA spectral alignment guidance | ADM · ADM ImageNet 256 | FID: 9.29 -> 7.81 | Improved |
| add SPA spectral alignment guidance | ADM · ADM ImageNet 256 | Density: 1.133 -> 1.226 | Improved |
| add SPA spectral alignment guidance | DDPM · DDPM CelebA | FID: 49.44 -> 48.63 | Improved |
| use eps-rescaling instead of SPA | DDPM · DDPM CelebA | FID: 48.63 -> 48.53 | Mixed |
| add time-shift sampling | ADM · ADM ImageNet 256 | FID: 9.29 -> 15.35 | Hurt |
| add SPA spectral alignment guidance | SDXL · SDXL COCO 5K (w=7.5) | HPSv3: 8.426 -> 8.829 | Improved |
| add SPA spectral alignment guidance | SD2.0 · SD2.0 COCO 5K (w=7.5) | HPSv3: 7.043 -> 7.238 | Improved |
| add SPA spectral alignment guidance | SD3.5 · SD3.5 COCO 5K (w=5) | HPSv3: 9.782 -> 9.975 | Improved |
| add SPA spectral alignment guidance | FLUX.1 dev · FLUX.1 dev COCO 5K (w=2.5) | win-rate: 53.3 +- 1.5% | Improved |
| add SPA spectral alignment guidance | FLUX.1 dev · FLUX.1 dev COCO 5K (w=3.5) | HPSv3: 12.47 -> 12.48 | No clear effect |
| add SPA guidance to FLUX.1 with true CFG de-distilled checkpoint | FLUX.1 dev · FLUX.1 dev COCO 5K (w=2.5) | HPSv3: 11.55 -> 11.76 | Improved |
| add SPA guidance | SDXL · SDXL | sampling time overhead: +3.86% | No clear effect |
Closest papers by meaning
- An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.87)
- Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps (ICLR 2024 · similarity 0.87)
- AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers (NeurIPS 2026 · similarity 0.87)
- Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models (NeurIPS 2026 · similarity 0.87)
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.86)
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.
