Research map / Image generation / Fast diffusion sampling
Learning Few-Step Diffusion Models by Trajectory Distribution Matching
Yihong Luo, Tianyang Hu, Jiacheng Sun, Yujun Cai, Jing Tang
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 unify training across step counts (ablation) | TDM-unify · HPS (SD-v1.5) 1 step | HPS: 20.81 -> 28.90 | Improved |
| add unify training across step counts (ablation) | TDM-unify · HPS (SD-v1.5) 2 steps | HPS: 29.08 -> 30.52 | Improved |
| add sampling-step conditioning (ablation) | TDM-unify · HPS (SD-v1.5) 1 step | HPS: 26.11 -> 28.90 | Improved |
| add sampling-step conditioning (ablation) | TDM-unify · HPS (SD-v1.5) 4 steps | HPS: 29.39 -> 31.31 | Improved |
| add surrogate training objective (ablation) | TDM-unify · HPS (SD-v1.5) 1 step | HPS: 28.23 -> 28.90 | Improved |
| add importance sampling for fake score (ablation) | TDM 4-step · HPS (SD-v1.5) | HPS: 29.27 -> 31.35 | Improved |
| match noisy samples instead of clean samples (ablation) | TDM 4-step · HPS (SD-v1.5) | HPS: 24.63 -> 31.35 | Improved |
| replace reverse KL by Fisher divergence (ablation) | TDM 4-step · HPS (SD-v1.5) | HPS: 31.35 -> 31.70 | Improved |
| use DPMSolver at test time instead of DDIM (ablation) | TDM 4-step · HPS (SD-v1.5) | HPS: 31.04 -> 31.35 | Improved |
| train with DPMSolver instead of DDIM (ablation) | TDM 4-step · HPS (SD-v1.5) | HPS: 31.35 -> 31.30 | No clear effect |
| fine-tune SD-v1.5 before distillation instead of GAN | TDM-unify · HPS (SD-v1.5) 4 steps | HPS: 30.83 -> 31.31 | Improved |
| use TDM over teacher diffusion | CogVideoX-2B · VBench | Total score: 80.91 -> 81.65 | Improved |
Closest papers by meaning
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.91)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.91)
- Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.91)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.91)
- T-Stitch: Accelerating Sampling in Pre-Trained Diffusion Models with Trajectory Stitching (ICLR 2025 · similarity 0.90)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · 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.
