Research map / Image generation / Text-to-image diffusion
DRM: Diffusion-based Reward Model With Step-wise Guidance
Jaxon Zhang, Binxin Yang, Hubery Yin, Chen Li, Jing LYU
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 |
|---|---|---|---|
| initialize DRM from pre-trained diffusion weights instead of random (ablation) | DRM · HPDv3 test preference accuracy | accuracy (%): 59.3 -> 74.0 | Improved |
| initialize from pre-trained weights rather than 3 epochs random training (ablation) | DRM · HPDv3 test preference accuracy | accuracy (%): 63.0 -> 71.9 | Improved |
| raise training resolution from 256 to 512 (ablation) | DRM · HPDv3 test preference accuracy | accuracy (%): 71.9 -> 74.0 | Improved |
| raise training resolution from 256 to 512 (ablation) | DRM · PickScore test preference accuracy | accuracy (%): 72.1 -> 73.4 | Improved |
| evaluate DRM on noisier latents (timestep 750 vs 0) (ablation) | DRM · HPSv3 test | accuracy (%): 74.0 -> 65.11 | Hurt |
| use Step-GRPO instead of standard GRPO with DRM (ablation) | DRM · RL fine-tuning, HPSv3 | HPSv3: 10.07 -> 10.28 | Improved |
| use Step-GRPO instead of standard GRPO with DRM (ablation) | DRM · RL fine-tuning, ImageReward | ImageReward: 1.14 -> 1.17 | Improved |
| use DRM instead of HPSv3 as reward under GRPO (ablation) | GRPO · RL fine-tuning, HPSv3 | HPSv3: 9.71 -> 10.07 | Improved |
| use DRM instead of HPSv3 as reward under GRPO (ablation) | GRPO · RL fine-tuning, ImageReward | ImageReward: 1.15 -> 1.14 | No clear effect |
| add Step-wise Sampling with k=6 candidates (ablation) | SD3.5-Medium · Step-wise Sampling, HPSv3 | HPSv3: 8.95 -> 9.49 | Improved |
| add Step-wise Sampling with k=6 candidates (time cost) (ablation) | SD3.5-Medium · Step-wise Sampling | time (s): 2.88 -> 9.83 | Hurt |
| increase Step-wise Sampling candidates from 2 to 6 (ablation) | SD3.5-Medium · Step-wise Sampling, HPSv3 | HPSv3: 9.02 -> 9.49 | Improved |
Closest papers by meaning
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.89)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.89)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.89)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.89)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · 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.
