Research map / Image generation / Fast diffusion sampling
Inference-Time Diffusion Model Distillation
Geon Yeong Park, Sang Wan Lee, Jong Chul Ye
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 inference-time distillation with teacher guidance (++) | LCM · MS-COCO 10K | FID: 20.674 -> 20.149 | Improved |
| add inference-time distillation with teacher guidance (++) | LCM-LoRA · MS-COCO 10K | FID: 20.300 -> 19.815 | Improved |
| add inference-time distillation with teacher guidance (++) | SDXL-Lightning · MS-COCO 10K | ImageReward: 0.787 -> 0.820 | Improved |
| add inference-time distillation with teacher guidance (++) | SDXL-Lightning LoRA · MS-COCO 10K | PickScore: 0.482 -> 0.518 | Improved |
| add inference-time distillation with teacher guidance (++) | DMD2 · MS-COCO 10K | FID: 21.238 -> 20.937 | Improved |
| add inference-time distillation with teacher guidance (++) | SDXL-Turbo · MS-COCO 10K | FID: 18.612 -> 18.481 | Improved |
| add one teacher step (4+1) instead of a fifth student step | LCM · MS-COCO 10K | FID: 20.732 -> 20.149 | Improved |
| add one teacher step (5+1) instead of a sixth student step | LCM · MS-COCO 10K | FID: 21.540 -> 20.369 | Improved |
| use renoising schedule s = t - dt instead of random timestep (ablation) | DMD2++ · MS-COCO 10K | FID: 21.105 -> 20.937 | Improved |
| use renoising schedule s = t - dt instead of s = t (ablation) | DMD2++ · MS-COCO 10K | FID: 21.342 -> 20.937 | Improved |
| use s = t instead of baseline DMD2 sampling (ablation) | DMD2++ renoising · MS-COCO 10K | ImageReward: 0.777 -> 0.777 | No clear effect |
| add the 1-step teacher correction | LCM 4-step sampling · MS-COCO 10K wall time | time (sec): 1.996 -> 1.987 | No clear effect |
Closest papers by meaning
- Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation (ICLR 2025 · similarity 0.91)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.90)
- Learnable Sampler Distillation for Discrete Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.90)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Data-free Distillation of Diffusion Models with Bootstrapping (ICML 2024 · similarity 0.89)
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.
