Research map / 3D vision / 3D generation
From Sparse to Dense: Multi-View GRPO for Flow Models via Augmented Condition Space
Jiazi Bu, Pengyang Ling, Yujie Zhou, Yibin Wang, Yuhang Zang, Tianyi Wei, Xiaohang Zhan, Jiaqi Wang, Tong Wu, Xingang Pan, Dahua Lin
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 augmented multi-view conditions (K=12, VLM enhancer) (ablation) | Flow-GRPO-Fast · HPD eval (HPS-v3 reward training) | HPS-v3: 0.149 -> 0.155 | Improved |
| add augmented multi-view conditions (K=12) (ablation) | Flow-GRPO-Fast · HPD eval (HPS-v3 reward training) | ImageReward: 1.152 -> 1.193 | Improved |
| add augmented multi-view conditions (K=12) (ablation) | Flow-GRPO-Fast · HPD eval (HPS-v3 reward training) | CLIP score: 0.378 -> 0.370 | Hurt |
| increase K from 6 to 12 (ablation) | MV-GRPO · HPD eval (HPS-v3 reward training) | HPS-v3: 0.154 -> 0.155 | Improved |
| derive conditions from distinct SDE samples (ablation) | MV-GRPO · HPD eval (HPS-v3 reward training) | HPS-v3: 0.153 -> 0.155 | Improved |
| add multi-perspective VLM prompt set (ablation) | MV-GRPO · HPD eval (HPS-v3 reward training) | HPS-v3: 0.152 -> 0.155 | Improved |
| use 8B instead of 2B VLM enhancer (ablation) | MV-GRPO · HPD eval (HPS-v3 reward training) | HPS-v3: 0.153 -> 0.155 | Improved |
| use online VLM instead of offline LLM enhancer | MV-GRPO · HPD eval (UnifiedReward-v2 reward training) | ImageReward: 1.118 -> 1.187 | Improved |
| MV-GRPO vs best baseline on HPS-v3 reward | MV-GRPO (VLM) · HPD eval (HPS-v3 reward training) | HPS-v3: 0.150 -> 0.155 (TempFlow-GRPO) | Improved |
| MV-GRPO vs baselines on CLIP score | MV-GRPO (VLM) · HPD eval (HPS-v3 reward training) | CLIP score: 0.380 -> 0.370 (Flow-GRPO) | Hurt |
| integrate MV-GRPO into DanceGRPO-Fast | DanceGRPO-Fast · HPD eval (HPS-v3 reward training) | HPS-v3: 0.149 -> 0.154 | Improved |
| use condition augmentation instead of data augmentation | Flow-GRPO-Fast · iteration time | iteration time (s): 1931.15 -> 191.95 | Improved |
Closest papers by meaning
- TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS (ICLR 2026 · similarity 0.92)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.90)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.87)
- MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge (NeurIPS 2026 · similarity 0.87)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · 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.
