Research map / Video / Video generation
VELR: Efficient Video Reward Feedback via Ensemble Latent Reward Models
Liyu Zhang, Kehan Li, Tao Zhou, Zeyi Huang, Chao Li, Jiming Chen
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 |
|---|---|---|---|
| replace full ReFL with latent reward model (VELR) | Wan-2.1 with UnifiedReward · memory during backprop | total memory (GB): 178.59 -> 22.18 | Improved |
| use ensemble of 10 LRMs instead of single LRM (ablation) | latent reward model · LRM test set | MSE: 0.851 -> 0.008 | Improved |
| use VELR LRM instead of Dollar LRM (ablation) | latent reward model · LRM test set | PCC: 0.34 -> 0.91 | Improved |
| fine-tune with VELR unified-reward | CogVideoX-1.5 5B · VBench2.0 Average | score: 52.25 -> 57.16 | Improved |
| fine-tune with VELR Causal-VQA | CogVideoX-1.5 5B · VBench2.0 Aesthetic Quality | score: 62.53 -> 61.94 | Hurt |
| start gradients at step 30 (late) instead of mid (ablation) | VELR · VBench Image Quality | score: 69.62 -> 66.39 | Hurt |
| shorten truncated duration to 4 (ablation) | VELR · VBench Image Quality | score: 69.62 -> 66.90 | Hurt |
| add online alignment of LRM (ablation) | VELR · VBench Image Quality | score: 68.36 -> 69.62 | Improved |
| add online alignment of LRM (ablation) | VELR · VBench Aesthetic Quality | score: 66.34 -> 66.21 | No clear effect |
| add uncertainty weighting (ablation) | VELR · VBench | score and GPU hours: GPU hours 15.0 -> 12.6 | Improved |
| add KL divergence term (ablation) | VELR · qualitative generation diversity | diversity | Improved |
| add truncated mid-step setting vs VADER/ImageReward paradigms (ablation) | LRM-based ReFL on Wan-2.1 · qualitative | video quality | Improved |
Closest papers by meaning
- T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback (NeurIPS 2024 · similarity 0.87)
- Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling (NeurIPS 2026 · similarity 0.87)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 2024 · similarity 0.86)
- Boosting Text-to-Video Generative Model with MLLMs Feedback (NeurIPS 2024 · similarity 0.85)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.85)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.85)
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.
