Research map / Image generation / Fast diffusion sampling
Efficient Diffusion Models via Time Step Optimization with Consistent Training and Inference Constraints
Binrui Wu, Zihao Cheng, Yuesen Liao, WEIZHONG ZHANG
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 |
|---|---|---|---|
| retrain the model only on the selected time steps | EDM diffusion model · CIFAR10 (NFE=18) | FID: 7.41 -> 6.15 | Improved |
| retrain the model only on the selected time steps | EDM diffusion model · AFHQv2 (NFE=18) | FID: 8.18 -> 6.99 | Improved |
| retrain on selected time steps | EDM diffusion model · CIFAR10 (NFE=256) | FID: 2.03 -> 1.84 | Improved |
| add fine-tuning on selected steps (PCS+ft) | PCS with DDIM · CIFAR10 (NFE=4) | FID: 25.22 -> 21.09 | Improved |
| add fine-tuning on selected steps (PCS+ft) | PCS with iPNDM · ImageNet 64x64 (NFE=4) | FID: 14.97 -> 11.75 | Improved |
| add fine-tuning on selected steps (PCS+ft) | PCS with iPNDM · AFHQv2 (NFE=4) | FID: 10.15 -> 8.98 | Improved |
| use PCS instead of LD3 schedule | DDIM · CIFAR10 (NFE=4) | FID: 26.04 -> 25.22 | Improved |
| use PCS instead of LD3 schedule | DDIM · AFHQv2 (NFE=4) | FID: 20.54 -> 21.32 | Hurt |
| use PCS instead of EDM handcrafted schedule | DDIM · CIFAR10 (NFE=4) | FID: 66.76 -> 25.22 | Improved |
| use PCS instead of EDM handcrafted schedule | iPNDM · CIFAR10 (NFE=4) | FID: 29.98 -> 8.96 | Improved |
| increase candidate set size from 32 to 128 (ablation) | PCS with DDIM · CIFAR10 (NFE=4) | FID: 29.76 -> 25.22 | Improved |
| increase candidate set size from 64 to 128 (ablation) | PCS with DDIM · CIFAR10 (NFE=4) | FID: 29.33 -> 25.22 | Improved |
Closest papers by meaning
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 2024 · similarity 0.93)
- Enhanced Diffusion Sampling via Extrapolation with Multiple ODE Solutions (ICLR 2025 · similarity 0.93)
- Accelerating Diffusion Sampling with Optimized Time Steps (CVPR 2024 · similarity 0.92)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.92)
- Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training (CVPR 2025 · similarity 0.91)
- Efficient Denoising Diffusion via Probabilistic Masking (ICML 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.
