Research map / Large language models / Alignment and preferences
Circular-DPO: Aligning Multi-Stage 3D Generative Models via Preference Feedback Loop
Zejian Li, Jiarui Ma, Han Xu, Weiting Zheng, Yangrui Zhu, Chenye Meng, Pei Chen, Ling Yang, Zhiyuan Yang, Changyuan Yang, Guang Yang, Immanuel Koh and 1 more
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 the data feedback loop that trains stage 1 on pairs from the aligned stage 2, versus independent stage-wise DPO (ablation) | Circular-DPO on Trellis · Test set of 100 prompts | ImageReward: -0.3663 -> -0.3146 | Improved |
| add the data feedback loop (versus independent stage-wise DPO) (ablation) | Circular-DPO on Trellis · Test set of 100 prompts | Reward3D: -0.5969 -> -0.5607 | Improved |
| add quality-gap filter and weight w1 (ablation) | Circular-DPO on Trellis · Test set of 100 prompts | Reward3D: -0.6199 -> -0.5607 | Improved |
| add quality-gap filter and weight w1 (ablation) | Circular-DPO on Trellis · Test set of 100 prompts | ImageReward: -0.3246 -> -0.3146 | Improved |
| add preference reliability weight w2 (ablation) | Circular-DPO on Trellis · Test set of 100 prompts | ImageReward: -0.4908 -> -0.3146 | Improved |
| use DPO instead of plain fine-tuning on stage 1 (ablation) | Trellis sparse structure model · Test set of 100 prompts | ImageReward: -0.4187 -> -0.3748 | Improved |
| use DPO instead of plain fine-tuning on stage 2 (ablation) | Trellis structured latent model · Test set of 100 prompts | ImageReward: -0.4410 -> -0.5512 | Hurt |
| apply DPO only to stage 2 (without stage 1 loop), vs untuned Trellis (ablation) | Trellis · Test set of 100 prompts | Reward3D: -0.7137 -> -0.7112 | No clear effect |
| apply Circular-DPO to Trellis | Trellis · Test set of 100 prompts | ImageReward: -0.4596 -> -0.3146 | Improved |
| apply Circular-DPO to MVDream | MVDream · Test set of 100 prompts | ImageReward: -0.3948 -> -0.3747 | Improved |
| compare Circular-DPO on Trellis against score-distillation DreamReward | Circular-DPO on Trellis · Test set of 100 prompts | Reward3D: -0.5607 vs 0.2733 | Hurt |
Closest papers by meaning
- DreamAlign: Dynamic Text-to-3D Optimization with Human Preference Alignment (AAAI 2025 · similarity 0.89)
- DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization (ICML 2025 · similarity 0.89)
- MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences (AAAI 2025 · similarity 0.87)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.86)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.86)
- Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (CVPR 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.
