Tencent Hunyuan open-sources EmoReward, a self-evolving reward calibration mechanism for emotional video generation RL
Model: HunyuanVideo-1.5 · availability, license and releases
Tencent Hunyuan released EmoReward, a self-evolving training mechanism that calibrates multi-reward weight allocation for emotional video-generation RL without updating reward models. It distills evidence reliability from reward traces and plugs into SAGE-GRPO on HunyuanVideo-1.5. The repository is available on GitHub.
China context
- Original name
- 腾讯混元
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the EmoReward code and integrate it into their own video generation RL pipelines, as the repository is publicly available on GitHub.
- For investors
- The open-source release of a training mechanism by Tencent Hunyuan indicates continued investment in video generation research, which may influence competitive dynamics in the AI video market.
EmoReward is a self-evolving training mechanism for emotional video-generation reinforcement learning. It distills evidence reliability from reward traces collected during a training run to guide multi-reward weight allocation in the next run, without updating any reward model. The mechanism includes Reflective Reward Gating for instance-level calibration and RewardMemory for cross-run experience reuse. It is designed as a drop-in reward for SAGE-GRPO training on HunyuanVideo-1.5 and transfers to DanceGRPO and FlowGRPO. The repository is open-sourced on GitHub under Tencent Hunyuan.
EmoReward introduces a Plan-Act-Reflect-Evolve pipeline: it grounds prompts into EmoScripts, scores videos with heterogeneous reward primitives, calibrates contributions via instance-level gates, and distills reliability into bounded fusion-weight multipliers for the next run. This addresses reward hacking and subjective evaluation in emotional video generation by making reward calibration adaptive without updating reward models.
Developers working on emotional video generation can now use EmoReward to improve RL training stability and output quality without retraining reward models, potentially reducing compute costs and iteration time. The open-source release may pressure competing labs to adopt similar self-evolving reward mechanisms.
For AI labs and video generation startups, EmoReward offers a way to improve emotional expressiveness in generated videos, which could differentiate products in entertainment, advertising, and virtual human applications. The open-source license allows commercial use, subject to the repository's license terms.
Next signals to watch: adoption of EmoReward in third-party video generation pipelines, benchmark results on public emotional video datasets, and whether Tencent Hunyuan integrates it into production models beyond HunyuanVideo-1.5.