Research map / Image generation / Text-to-image diffusion
DenseGRPO: From Sparse to Dense Reward for Flow Matching Model Alignment
Haoyou Deng, Keyu Yan, Chaojie Mao, Xiang Wang, Yu Liu, Changxin Gao, Nong Sang
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 ODE-based step-wise dense reward (DenseGRPO) over Flow-GRPO | Flow-GRPO · Human Preference Alignment | PickScore: 23.31 -> 24.64 | Improved |
| add DenseGRPO over Flow-GRPO | Flow-GRPO · Compositional Image Generation | GenEval: 0.95 -> 0.97 | Improved |
| add DenseGRPO over Flow-GRPO | Flow-GRPO · Visual Text Rendering | OCR accuracy: 0.92 -> 0.95 | Improved |
| use ODE-based dense reward instead of latent-similarity (CoCA) dense reward | Flow-GRPO · Human Preference Alignment | PickScore: 23.63 -> 24.64 | Improved |
| add latent-similarity step reward (CoCA) to Flow-GRPO | Flow-GRPO · Human Preference Alignment | PickScore: 23.31 -> 23.63 | Improved |
| use dense reward R_{t-1} per step instead of sparse terminal reward (ablation) | Flow-GRPO · Human Preference Alignment | PickScore | Improved |
| use calibrated timestep-specific noise level instead of uniform a=0.7 (ablation) | DenseGRPO · Human Preference Alignment | PickScore | Improved |
| use uniform a=0.7 noise with dense reward instead of Flow-GRPO (ablation) | DenseGRPO · Human Preference Alignment | PickScore | Improved |
| increase ODE denoising steps from n=1 to n=t (ablation) | DenseGRPO · Human Preference Alignment | PickScore | Improved |
| use single-step ODE (n=1) for dense reward instead of Flow-GRPO sparse reward (ablation) | DenseGRPO · Human Preference Alignment | PickScore | Hurt |
| increase ODE steps n from 1 to 2 to t (ablation) | DenseGRPO · training cost | GPU hours for 20 steps: 11, 13, and 19 GPU hours | Hurt |
| apply DenseGRPO on FLUX.1-dev | Flow-GRPO · FLUX.1-dev | PickScore | Improved |
Closest papers by meaning
- Alleviating Sparse Rewards by Modeling Step-Wise and Long-Term Sampling Effects in Flow-Based GRPO (ICML 2026 · similarity 0.91)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.89)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.87)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.87)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.86)
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.86)
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.
