Research map / Image generation / Fast diffusion sampling
On the Design of One-step Diffusion via Shortcutting Flow Paths
Haitao Lin, Peiyan Hu, Minsi Ren, Zhifeng Gao, Zhi-Ming Ma, Guolin Ke, Tailin Wu, Stan Z. Li
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 |
|---|---|---|---|
| use ESC improvements instead of MeanFlow baseline (240 epochs, 1-NFE) | MeanFlow SiT-XL/2 · ImageNet-256 (1-NFE, CFG) | FID50k: 3.43 -> 2.85 | Improved |
| train ESC for 480 epochs instead of 240 | ESC SiT-XL/2 · ImageNet-256 (1-NFE, CFG) | FID50k: 2.85 -> 2.53 | Improved |
| use plug-in velocity (minibatch mixture) instead of conditional velocity (ablation) | continuous-time shortcut model · ImageNet-256 training cost | time per iteration (ms): 554 -> 558 ms/iter | No clear effect |
| add class-consistent mini-batching under CFG training (ablation) | ESC · ImageNet-256 | FID50k final | No clear effect |
| add class-consistent mini-batching under CFG training (ablation) | ESC · ImageNet-256 | FID50k convergence speed | Improved |
| use linear paths instead of cosine paths (ablation) | CT and sCT shortcut models · CIFAR-10 / ImageNet-256 | FID50k | Improved |
| use continuous-time instead of discrete-time shortcut models (ablation) | CT and SCD · CIFAR-10 (unconditional) | FID50k | Improved |
| sample terminal time r randomly instead of fixing r=0 (ablation) | sCT-linear vs MeanFlow · CIFAR-10 / ImageNet-256 | FID50k | Mixed |
| add gradual time sampler from sCT to MeanFlow, K_fix0 about 20k (ablation) | MeanFlow · ImageNet-256 SiT-B/2 | FID50k | Improved |
| add variational adaptive loss weighting and tangent warmup (ablation) | continuous-time shortcut model · CIFAR-10 / ImageNet-256 | FID50k | Improved |
Closest papers by meaning
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.92)
- Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction (NeurIPS 2025 · similarity 0.90)
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.90)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.89)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · 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.
