Research map / Image generation / Fast diffusion sampling
Rdm: Re-conceptualizing Distribution Matching as a Reward for Diffusion Distillation
Linqian Fan, Peiqin Sun, Tiancheng Wen, Shun Lu, Chengru Song
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 group normalization to the distribution matching reward (ablation) | GNDMR · SD3-Medium 512, only Rdm | FID: 24.94 -> 23.07 | Improved |
| add group normalization with HPS reward (ablation) | GNDMR · SD3-Medium 512 +HPS | FID: 25.40 -> 24.47 | Improved |
| add group normalization with HPS reward (ablation) | GNDMR · SD3-Medium 512 +HPS | HPS: 30.39 -> 30.22 | No clear effect |
| add group normalization with PS reward (ablation) | GNDMR · SD3-Medium 512 +PS | FID: 24.61 -> 22.32 | Improved |
| share diffused timestep t' within group (ablation) | GNDMR · SD3-Medium 512 | FID: 24.94 -> 24.34 | Improved |
| additionally share generator timestep t within group (ablation) | GNDMR · SD3-Medium 512 | FID: 24.34 -> 23.07 | Improved |
| apply group normalization over larger diffused timestep interval (ablation) | GNDMR · SD3-Medium 512 | FID: 23.71 -> 23.43 | Improved |
| share initial noise within a group instead of random (ablation) | GNDMR · SD3-Medium 512 | FID: 24.68 -> 23.07 | Improved |
| use sample-wise adaptive weight beta_dm over fixed weight (ablation) | GNDMR · SD3-Medium 512 HPS | HPS | Improved |
| use importance sampling for multi-step updates (GNDMR-IS) | GNDMR · SD3-Medium 1024 | HPS: 30.37 -> 30.00 | Mixed |
| use GNDMR instead of DMDR | DMDR · SD3-Medium 1024 | FID: 29.10 -> 28.02 | Improved |
| use GNDMR instead of DMDR | DMDR · SD3.5-Medium 512 | HPS: 30.83 -> 31.25 | Improved |
Closest papers by meaning
- GDMD: Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.92)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.90)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.90)
- Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models (NeurIPS 2026 · similarity 0.89)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.89)
- Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation (ICLR 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.
