Research map / Image generation / Fast diffusion sampling
Score Distillation Beyond Acceleration: Generative Modeling from Corrupted Data
Yasi Zhang, Tianyu Chen, Zhendong Wang, Ying Nian Wu, Mingyuan Zhou, Oscar Leong
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 |
|---|---|---|---|
| distill corruption-aware teacher into one-step generator (RSD) | Teacher-Truncated · CIFAR-10 denoising (sigma=0.2) | FID: 12.21 -> 4.77 | Improved |
| distill teacher diffusion into one-step generator | Teacher Diffusion · CelebA-HQ random inpainting (p=0.9) | FID: 25.53 -> 16.86 | Improved |
| distill teacher diffusion into one-step generator | Teacher Diffusion · CelebA-HQ Gaussian deblurring (sigma=0.2) | FID: 99.19 -> 76.98 | Improved |
| distill Fourier-space ambient diffusion teacher | Teacher Diffusion · Multi-coil MRI R=8 | FID: 48.15 -> 22.51 | Improved |
| distill into one-step generator | Teacher-Truncated · CIFAR-10 denoising (sigma=0.4) | FID: 22.12 -> 21.63 | No clear effect |
| reduce training set to 10% of CelebA-HQ (ablation) | RSD · CelebA-HQ denoising | FID: 6.48 -> 10.53 | Hurt |
| use SiD distillation loss instead of SDS or DMD (ablation) | RSD · multiple datasets | FID | Improved |
| use different objectives for pretraining and fake-model training (ablation) | RSD · training stability | training stability | Hurt |
| apply corruption pipeline to generator outputs for distillation | RSD · denoising and restoration | FID | Improved |
| one-step generator replaces multi-step teacher sampling | Teacher diffusion · CIFAR-10 sigma=0.2, 50,000 samples | sampling time: 10 minutes -> 20 seconds | Improved |
| select checkpoints with Proximal FID instead of oracle true FID (ablation) | RSD · CIFAR-10 denoising | FID: 4.76 -> 5.21 | No clear effect |
Closest papers by meaning
- Noise Conditional Variational Score Distillation (ICML 2025 · similarity 0.90)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.90)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.89)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.89)
- Data-free Distillation of Diffusion Models with Bootstrapping (ICML 2024 · similarity 0.89)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · 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.
