Research map / Image generation / Fast diffusion sampling
DistillKac: Few-Step Image Generation via Damped Wave Equations
Weiqiao Han, Chenlin Meng, Christopher D Manning, Stefano Ermon
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 endpoint-only distillation to Guided Kac Flow (CelebA-64, 1 NFE) (ablation) | Guided Kac Flow · CelebA-64 (1 NFE) | FID: 443.01 -> 7.45 | Improved |
| add endpoint-only distillation to Guided Kac Flow (CelebA-64, 20 NFE) (ablation) | Guided Kac Flow · CelebA-64 (20 NFE) | FID: 11.23 -> 3.42 | Improved |
| distill the teacher to 1 step on CIFAR-10 (ablation) | Guided Kac Flow · CIFAR-10 (1 step) | FID: 396.46 -> 5.66 | Improved |
| use multi-stage progressive distillation instead of single-stage 100 -> 1 (ablation) | DistillKac · CIFAR-10 (1 step, w=1.2) | FID: 9.71 -> 5.66 | Improved |
| use multi-stage 4 -> 2 -> 1 instead of 4 -> 1 (ablation) | DistillKac · CIFAR-10 (1 step, w=1.2) | FID: 5.72 -> 5.66 | No clear effect |
| use second-order integrator (midpoint) instead of Euler for 100-step Guided Kac Flow (ablation) | Guided Kac Flow · CIFAR-10 (w=1.2) | FID: 4.00 -> 3.54 | Improved |
| use AB-2 instead of midpoint integrator (ablation) | Guided Kac Flow · CIFAR-10 (w=1.2) | FID: 3.54 -> 3.58 | No clear effect |
| add velocity-space classifier-free guidance (tune w to 1.2) (ablation) | Guided Kac Flow · CIFAR-10 | FID: 4.20 -> 3.54 (w=1.0 -> 1.2, midpoint) | Improved |
| increase guidance strength beyond w=1.2 (ablation) | Guided Kac Flow · CIFAR-10 | FID: 3.54 -> 4.10 (w=1.2 -> 1.5) | Hurt |
| use mean-reverting Kac hyperparameters a=3000, c=20 with g(t)=t^2 (ablation) | Kac flow · CelebA-64 lightweight UNet (2k samples) | FID: 19.38 -> 15.30 | Improved |
| increase damping a from 3000 to 6000 at c=20 (ablation) | Kac flow · CelebA-64 lightweight UNet (2k samples) | FID: 15.30 -> 16.96 | Hurt |
| reduce student step count from 20 to 1 under endpoint distillation (ablation) | DistillKac · CIFAR-10 | FID: 3.72 -> 5.66 | Hurt |
Closest papers by meaning
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.86)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.85)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.85)
- D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models (NeurIPS 2026 · similarity 0.85)
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.
