Research map / Image generation / Fast diffusion sampling
Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis
Tianhe Wu, Ruibin Li, Lei Zhang, Kede Ma
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 first-step teacher-derived v-prediction diversity loss with detached DMD gradient | DMD (SD3.5-M, 4 NFEs) · Pick-a-Pic | DINO diversity: 0.137 -> 0.179 | Improved |
| add first-step diversity loss with detach | DMD (SDXL, 4 NFEs) · Pick-a-Pic | DINO diversity: 0.109 -> 0.173 | Improved |
| add first-step diversity loss with detach | DMD (SD3.5-M, 4 NFEs) · Pick-a-Pic | ImageReward: 1.189 -> 1.142 | Hurt |
| add LPIPS perceptual regularization | DMD (SD3.5-M) · Pick-a-Pic | DINO diversity: 0.137 -> 0.169 | Mixed |
| add adversarial discriminator regularization | DMD (SD3.5-M) · Pick-a-Pic | VQ-R1 quality: 4.649 -> 4.525 | Hurt |
| move diversity anchor step K from 1 to 5 (ablation) | DP-DMD · Pick-a-Pic | DINO diversity: 0.157 -> 0.179 | Improved |
| move anchor step K from 5 to 10 (ablation) | DP-DMD · Pick-a-Pic | VQ-R1 quality: 4.646 -> 4.589 | Hurt |
| increase diversity weight lambda to 0.10 (ablation) | DP-DMD · Pick-a-Pic | ImageReward: 1.125 -> 1.056 | Hurt |
| stop gradients after the first step (detach) (ablation) | DP-DMD · Pick-a-Pic training curves | diversity | Improved |
| replace DMD with DP-DMD for prompt following | SD3.5-M distilled · GenEval | overall accuracy: 0.66 -> 0.65 | No clear effect |
| compare DP-DMD to TDM at 4 NFEs | DP-DMD · Pick-a-Pic | DINO diversity: 0.148 -> 0.162 | Improved |
Closest papers by meaning
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.94)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.89)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.88)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.88)
- DELT: A Simple Diversity-driven EarlyLate Training for Dataset Distillation (CVPR 2025 · similarity 0.88)
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.87)
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.
