Research map / Image generation / Fast diffusion sampling
Overclocking Electrostatic Generative Models
Daniil Shlenskii, Alexander Korotin
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 with IPFM (alpha=0.5) to 4 steps | PFGM++ D=128 · CIFAR-10 (unconditional) | FID: 1.92 (35 NFE teacher) vs 2.08 | No clear effect |
| distill with IPFM (alpha=0.5) to 2 steps | PFGM++ D=128 · FFHQ 64x64 | FID: 2.43 -> 2.12 | Improved |
| add SiD-inspired regularization (alpha 0.5 -> 1.0) (ablation) | IPFM D=128 · CIFAR-10, 4 steps | FID: 2.08 -> 1.75 | Improved |
| add SiD-inspired regularization (alpha 0.5 -> 1.0) (ablation) | IPFM D=128 · CIFAR-10, 1 step | FID: 5.38 -> 3.31 | Improved |
| add SiD-inspired regularization (alpha 0.5 -> 1.0) (ablation) | IPFM D=128 · FFHQ 64x64, 2 steps | FID: 2.12 -> 1.72 | Improved |
| add SiD-inspired regularization (alpha 0.5 -> 1.0) (ablation) | IPFM D=infinity · CIFAR-10, 4 steps | FID: 2.13 -> 1.86 | Improved |
| use finite D=128 instead of diffusion limit | IPFM one-step · CIFAR-10, 1 step, 70h | FID: 3.611 -> 3.216 | Improved |
| use SiD with heavy-tailed kernel instead of Gaussian (ablation) | SiD D=128 · CIFAR-10, 1 step, 70h | FID: 3.611 -> 3.336 | Improved |
| use full IPFM D=128 instead of SiD with heavy tails (ablation) | IPFM · CIFAR-10, 1 step, 70h | FID: 3.336 -> 3.216 | Improved |
| use IPFM instead of Consistency Trajectory Models (no GAN loss) | distillation · PFGM++ at D=128 and D=inf | FID | Improved |
| use finite D teacher (128 vs 2048) | IPFM alpha=1.0 · CIFAR-10, 4 steps | FID: 1.75 vs 1.82 | No clear effect |
Closest papers by meaning
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.90)
- EM Distillation for One-step Diffusion Models (NeurIPS 2024 · similarity 0.89)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.88)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.87)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.87)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.87)
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.
