Tencent's Adaptive Reward Routing Dynamically Balances Multi-Reward RL for Joint Audio-Video Diffusion
tencent · hf · 2026-10-02
Tencent proposes Adaptive Reward Routing for forward-process RL (DiffusionNFT) of joint audio-video diffusion models, dynamically adapting where reward updates act and how competing rewards combine. Cross-Modal Influence-Guided Routing uses bidirectional cross-attention responses as a proxy to reweight token-level losses and scale gradients across cross-modal layers; Preference-Preserving Modality-Aware Reweighting keeps predefined weights as priors with residual corrections from branch-level reward-gradient interactions. Experiments show consistent gains in modality quality, semantic consistency, and audio-video synchronization over strong RL baselines.
More from Multimodal
- A 3-pass workflow for restoring old photos with AI: repair, colorize, then sharpen — xiaohu · 2026-10-02
- Netflix's Align Then Reason lip-sync judge boosts mean AUC by up to 59% — netflix · 2026-10-02
- PROWBench: 170 Programmable Episodes Test Whether Video World Models Render What the Program Specifies — AlayaLab · 2026-10-02
- Stability AI's 4Director Controls Video World Models with Rigid 3D Geometry — stabilityai · 2026-10-02
- PixelDense: Dense-Prediction Teachers Beat Semantic Encoders for Diffusion REPA, GenEval Hits 0.8093 — Lehan Yang · 2026-10-02
- Opus 5.5 artwork sparks GPT-Image-2 vs Z-Image style comparison — teortaxesTex · 2026-10-02