EmoReward stops emotional video RL from trusting fixed rewards: last run's traces set the next run's weights

Reinforcement-learning post-training for emotional video generation now treats reward weights as experience that accumulates across training runs, rather than reusing one fixed fusion for every scene. Existing video GRPO methods typically rely on fixed rewards for visual quality, motion quality and text alignment that poorly capture emotional intent, and fixed fusion can mix informative, missing and saturated signals. On HunyuanVideo 1.5, within-optimizer comparisons show VideoAlign Overall rising from -1.2812 to -0.9594 for DanceGRPO with Reflect, from -0.7719 to -0.5376 for FlowGRPO with EmoReward, and from -0.8269 to -0.2624 for SAGE-GRPO with EmoReward.

The system first turns an emotional prompt into assessable cues, then allocates weight across scoring channels according to the validity of evidence in the current video. Plan converts the prompt into an EmoScript recording valence and arousal targets, facial action-unit intensities, body cues and an expression window; Act scores the video with seven fusion providers and keeps criterion scores separate from evidence validity. Reflect builds instance-level gates from training phase, face confidence, semantic conflict and subtlety, then renormalizes over active providers; a missing face channel's neutral default is excluded through its validity mask rather than entering the fusion.

After a completed training run, each provider's agreement with outcomes is compressed into bounded multipliers that the next run looks up directly. RewardMemory groups traces by emotional regime, estimates reliability from the correlation between raw provider scores and VideoAlign outcomes, and converts it into capped weight multipliers while reward-model parameters stay fixed during training. In a controlled comparison of seven reward policies at fixed backbone and optimization budget, EmoReward with RewardMemory reaches the highest reliability selectivity at 0.428 and both composite scores, core 0.922 and emo 0.884.

Evaluation centers on emotional video generation: HunyuanVideo 1.5 with DanceGRPO, FlowGRPO and SAGE-GRPO, averaged over 302 emotional prompts that do not overlap the training set; image transfer covers SD3.5 and FLUX, and human ratings come from 18 participants scoring three methods on 30 prompts.

Reliability is defined as a provider's agreement with VideoAlign outcomes, and VideoAlign is also an optimization objective, so that agreement is not independent emotional fidelity; the Macro channel shares evaluator information with the reference outcome. Sustained improvement across many generations is not evaluated, and among 27 memory hyperparameter settings only the final window exceeds fixed weighting in 23, with those settings reusing the same traces rather than independent training seeds.