Research map / Image generation / Fast diffusion sampling
Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield
Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Steven HOI, Hongsheng 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 |
|---|---|---|---|
| decouple re-noising schedules but keep both global (ablation) | DMD · DPG-Bench (Lumina-Image-2.0) | DPG overall: 83.90 -> 83.77 | No clear effect |
| constrain CA re-noising to tau > t (ablation) | DMD (Lumina-Image-2.0) · DPG-Bench | DPG overall: 83.77 -> 85.64 | Improved |
| keep DM regularizer over full noise range while constraining CA (hybrid) (ablation) | Decoupled-Constrained · HPS v2.1 | HPS v2.1 overall: 31.71 -> 32.29 | Improved |
| keep DM regularizer over full noise range while constraining CA (hybrid) (ablation) | Decoupled-Constrained · HPS v3 | HPS v3 overall: 11.08 -> 11.59 | Improved |
| replace DMD2 re-noising schedule with decoupled-hybrid | DMD2 · 4-step SDXL | FID: 18.95 -> 17.80 | Improved |
| replace DMD2 re-noising schedule with decoupled-hybrid | DMD2 · 4-step SDXL | HPS v2.1: 30.64 -> 30.34 | Hurt |
| replace DMD2 re-noising schedule with decoupled-hybrid | DMD2 · 4-step SDXL | ImageReward: 71.01 -> 78.61 | Improved |
| train with CFG Augmentation only (no DM) (ablation) | DMD · 4-step SDXL | stability / image quality | Mixed |
| replace DM with mean-variance KL regularizer (ablation) | CA engine · 4-step SDXL | stability / quality | Mixed |
| replace DM with GAN regularizer (ablation) | CA engine · 4-step SDXL | training stability | Hurt |
| restrict CA re-noising to clean timesteps [0.7, 1.0] (ablation) | CA-only single-step generator · image quality | training stability | Hurt |
Closest papers by meaning
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation (ICLR 2025 · similarity 0.88)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.88)
- GDMD: Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.87)
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 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.
