Research map / Image generation / Fast diffusion sampling
Progressive Distillation for Fast Sampling of Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Distill a deterministic DDIM teacher progressively, halving the sampling steps | Diffusion models · CIFAR-10 | FID: 3.00 | Helped |
| Use progressive distillation instead of knowledge distillation | CIFAR-10 diffusion sampling · CIFAR-10, 1 model evaluation | FID: 9.36 -> 9.12 | Helped |
| Use progressive distillation instead of DDIM | CIFAR-10 diffusion sampling · CIFAR-10, few sampling evaluations | FID: 13.36 -> 9.12 | Helped |
| Use direct x prediction with truncated SNR weighting rather than epsilon prediction with SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | FID: 2.91 -> 2.51 | Helped |
| Use combined x and epsilon prediction with truncated SNR weighting instead of SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | FID: 2.78 -> 2.76 | No clear effect |
| Use combined x and epsilon prediction with SNR+1 weighting instead of SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | FID: 2.78 -> 2.87 | No clear effect |
| Use epsilon prediction with truncated SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | training stability | Hurt |
| Use direct x prediction with truncated SNR weighting rather than direct x prediction with SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | FID: 2.75 -> 2.51 | Helped |
| Use v prediction with truncated SNR weighting rather than v prediction with SNR weighting (ablation) | Undistilled diffusion model · CIFAR-10 | FID: 3.05 -> 2.75 | Helped |
| Reduce training updates per student from the proposed schedule (ablation) | Progressive Distillation · CIFAR-10 | FID | Mixed |
| Halve sampling steps by a factor of 4 rather than 2 per distillation round (ablation) | Progressive Distillation · CIFAR-10 | FID | Hurt |
Closest papers by meaning
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.95)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.93)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Neural Diffusion Models (ICML 2024 · similarity 0.91)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.91)
- Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.91)
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.