Research map / Image generation / Fast diffusion sampling
Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis
Yanzuo Lu, Yuxi Ren, Xin Xia, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Andy J. Ma, Xiaohua Xie, Jian-Huang Lai
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 ADM fine-tuning after ADP pre-training (ablation) | DMDX one-step SDXL · COCO 10K (SDXL ablation) | HPSv2: 27.3499 -> 27.7046 | Improved |
| add ADM fine-tuning after ADP pre-training (ablation) | DMDX one-step SDXL · COCO 10K (SDXL ablation) | MPS: 10.6646 -> 11.1978 | Improved |
| add ADP pre-training before ADM (ablation) | ADM one-step SDXL · COCO 10K (SDXL ablation) | HPSv2: 26.8732 -> 27.7046 | Improved |
| replace DMD loss with ADM (both after ADP) (ablation) | DMDX one-step SDXL · COCO 10K (SDXL ablation) | HPSv2: 27.3985 -> 27.7046 | Improved |
| replace DMD loss with ADM (both without ADP) (ablation) | one-step SDXL w/o ADP · COCO 10K (SDXL ablation) | HPSv2: 25.9680 -> 26.8732 | Improved |
| use SAM instead of DINOv2 as pixel-space discriminator (ablation) | ADP pre-training · COCO 10K (SDXL ablation) | HPSv2: 27.1039 -> 27.3499 | Improved |
| add pixel-space discriminator weight lambda2=0.15 (vs 0.0) (ablation) | ADP pre-training · COCO 10K (SDXL ablation) | HPSv2: 26.7869 -> 27.3499 | Improved |
| lower pixel-space weight from 0.3 to 0.15 (ablation) | ADP pre-training · COCO 10K (SDXL ablation) | HPSv2: 26.8902 -> 27.3499 | Improved |
| ADP adversarial pre-training instead of single Rectified Flow reflow (ablation) | one-step SDXL pre-training · COCO 10K (SDXL ablation) | HPSv2: 23.6093 -> 27.3499 | Improved |
| increase TTUR from 1 to 8 (ablation) | DMDX one-step SDXL · COCO 10K | HPSv2: 27.7046 -> 27.7586 with training time x1.00 -> x2.53 | No clear effect |
| train 2x longer (ablation) | ADM 8-step CogVideoX-2b · VBench | final score: 78.584 -> 80.764 | Improved |
| add CFG to the few-step generator (ablation) | ADM 8-step CogVideoX-2b · VBench | semantic score: 69.621 -> 75.569 | Improved |
Closest papers by meaning
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.92)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.90)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.89)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.89)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Taming Diffusion for Dataset Distillation with High Representativeness (ICML 2025 · similarity 0.88)
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.
