Research map / 3D vision / 3D generation
Refining Few-Step Text-to-Multiview Diffusion via Reinforcement Learning
Ziyi Zhang, Li Shen, Deheng Ye, Yong Luo, Huangxuan Zhao, Meng Liu, Wei Yu, Lefei Zhang
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 advantage learning with zigzag sampling (MV-ZigAL) versus MV-RDL | MV-Adapter (LCM-SDXL) · Training prompts | HyperScore overall: 9.03 -> 9.17 | Improved |
| add advantage learning with zigzag sampling (MV-ZigAL) versus MV-PG | MV-Adapter (LCM-SDXL) · Training prompts | HyperScore overall: 8.39 -> 9.17 | Improved |
| use advantage learning instead of standard policy gradient with zigzag pass (ablation) | MVC-ZigAL · MATE-3D | HyperScore overall: 6.86 -> 7.04 | Improved |
| use first-step zigzag instead of full-step zigzag (ablation) | MVC-ZigAL · MATE-3D | HyperScore overall: 6.91 -> 7.04 | Improved |
| use multiview-constrained optimization instead of weighted sum (w_mv=0.1) (ablation) | MVC-ZigAL · MATE-3D | HyperScore overall: 6.25 -> 7.04 | Improved |
| use multiview-constrained optimization instead of weighted sum (w_mv=0.5) (ablation) | MVC-ZigAL · MATE-3D | HyperScore overall: 6.83 -> 7.04 | Improved |
| use multiview-constrained optimization instead of weighted sum (w_mv=1.0) (ablation) | MVC-ZigAL · MATE-3D | HyperScore overall: 6.68 -> 7.04 | Improved |
| use adaptive threshold instead of fixed thresholds (7.5 or 9.0) (ablation) | MVC-ZigAL · training reward curves | single-view and joint-view rewards: 7.5 and 9.0 fixed | Improved |
| use adaptive step size instead of fixed step sizes (0.01 or 0.1) (ablation) | MVC-ZigAL · training reward curves | constraint satisfaction: 0.01 and 0.1 fixed | Improved |
| scale RL finetuning to MATE-3D prompt set | MVC-ZigAL · MATE-3D | ImageReward: -0.846 -> 0.865 | Improved |
| add single-view-only reward (ZigAL) or weighted sum (ablation) | ZigAL / WS-ZigAL · Training prompts | joint vs per-view reward | Hurt |
Closest papers by meaning
- Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning (CVPR 2024 · similarity 0.90)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.89)
- Self-Play Fine-tuning of Diffusion Models for Text-to-image Generation (NeurIPS 2024 · similarity 0.89)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.89)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.88)
- Vivid-ZOO: Multi-View Video Generation with Diffusion Model (NeurIPS 2024 · similarity 0.88)
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.
