Research map / Video / Video generation
Aligning Human Sense: Calibrated Distributional Reward Learning for Video Generation
Naixin Zhai, Weihua Cheng, Dexu Yu, Yikai Gu, Hanwen Du, Junchen Fu, Chenxi Huang, Yingwei Song, Liyuan Ma, Yang Ran, Youhua Li, Yongxin Ni
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 |
|---|---|---|---|
| train reward model on calibrated full data instead of raw (ablation) | reward model · Reward model, Motion Effects (w/ ties) | pairwise accuracy: 65.66 -> 71.72 | Improved |
| train reward model on calibrated full data instead of raw (ablation) | reward model · Reward model, Action (w/o ties) | pairwise accuracy: 80.30 -> 83.33 | Improved |
| curate equal-scale data instead of raw (ablation) | reward model · Reward model, Action (w/ ties) | pairwise accuracy: 54.65 -> 59.60 | Improved |
| add Wasserstein loss to reward model (ablation) | reward model · Reward model, Motion Effects (w/ ties) | pairwise accuracy: 68.33 -> 71.72 | Improved |
| add Wasserstein loss to reward model (ablation) | reward model · Reward model, Action (w/ ties) | pairwise accuracy: 58.98 -> 60.61 | Improved |
| replace KL with Wasserstein in GRPO (ablation) | WGRPO · Video-Bench generation | avg score: 68.02 -> 69.58 | Improved |
| add KL regularizer to GRPO (ablation) | GRPO · Video-Bench generation | avg score: 67.97 -> 68.02 | No clear effect |
| KL instead of Wasserstein on dynamic degree (ablation) | GRPO · Video-Bench generation | dynamic degree: 36.50 vs 36.48 | Mixed |
| calibrated reward model for video generation | Wan2.1-T2V-1.3B · Video-Bench generation | avg score: 68.16 -> 69.58 | Improved |
| calibrated reward model for video generation | CogVideoX-2B · Video-Bench generation | subject consistency: 87.35 -> 86.45 | Hurt |
| calibrated reward model instead of raw | DanceGRPO · Video-Bench generation (optimizer comparison) | avg score: 68.99 -> 69.97 | Improved |
| calibrated reward model instead of raw | FlowGRPO · Video-Bench generation (optimizer comparison) | avg score: 69.86 -> 70.87 | Improved |
Closest papers by meaning
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.92)
- SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset (NeurIPS 2024 · similarity 0.89)
- Video-Bench: Human-Aligned Video Generation Benchmark (CVPR 2025 · similarity 0.87)
- Boosting Text-to-Video Generative Model with MLLMs Feedback (NeurIPS 2024 · similarity 0.87)
- Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria (NeurIPS 2026 · similarity 0.87)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 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.
