EmoReward self-evolving reward calibration advances emotional video RL
Original title: EmoReward: Self-Evolving Reward Calibration for Emotional Video Generation RL
EmoReward distills reliability from training traces to adaptively allocate weights for facial motion and chromatic cues in video generation.
Group Relative Policy Optimization helps generative models but struggles with subjective emotional video quality. EmoReward plans grounding cues then scores videos using heterogeneous signals with reflective gating. RewardMemory estimates primitive reliability from rollout outcomes to calibrate fusion weights in future runs. Experiments on HunyuanVideo 1.5 and transfers to SD 3.5 FLUX show gains across RL algorithms.
- Plans emotional grounding cues from prompts
- Applies instance-level gates for phase face conflict subtlety
- Estimates reliability via RewardMemory without updating reward models
- Improves multiple RL algorithms on HunyuanVideo 1.5
From the original article
1 min read at huggingface.co
Group Relative Policy Optimization (GRPO) has advanced generative model post-training in verifiable domains such as code and mathematics, but evaluating video quality, especially emotional expression, often depends on subjective aesthetics and has no uniquely correct target.
By Mingzhe Zheng, Yue Wu, Jing Tan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.