Tencent Hunyuan's EmoReward Self-Evolves Reward Weights for Video RL
Tencent-Hunyuan · hf · 2026-11-05
Tencent Hunyuan and HKUST introduce EmoReward, a self-evolving reward calibration mechanism for emotional video generation RL. It plans grounding cues from emotional prompts, scores videos with heterogeneous signal primitives, and uses Reflective Reward Gating for instance-level weight calibration. RewardMemory estimates each primitive's reliability from prior training rollouts to fix bounded multipliers for the next run—no reward model updates needed. Improvements shown on HunyuanVideo 1.5 across RL algorithms, with transfer to SD 3.5 and FLUX.
More from Multimodal
- Viggle Remake launches open-weight video transformation agent at $0.96/10s — Div_pradeep · 2026-11-05
- Soniox launches STT v6 with agent text input, first STT aware of both conversation sides — solyarisoftware · 2026-11-05
- Midjourney's --character design flag for consistent character concepts — hewarsaber · 2026-11-05
- Stunning 9-scene AI character consistency grid wows with same vampire heroine — flowersslop · 2026-11-05
- Creator puts Kling 4.0 vs 4.0 Flash to the test: control for polish, Flash for speed — techhalla · 2026-11-05
- TurnWise: Turn-Level Credit Assignment Advances Streaming Video RL — Shuai Dong · 2026-11-05