Stabilizing On-Policy Distillation via Critic Regularization
Stabilize on-policy distillation by applying critic regularization techniques like Spectral Normalization and target networks to prevent generator collapse.
Source paper
Ethan Chern, Zhulin Hu, Bohao Tang, Jiadi Su, Steffi Chern, Zhijie Deng, Pengfei Liu
Generated by AI. The paper’s authors did not write these.
Stabilize on-policy distillation by applying critic regularization techniques like Spectral Normalization and target networks to prevent generator collapse.
Replace fixed attention sinks with a learnable token eviction policy that dynamically evaluates the semantic importance of past key-value pairs.
Modify the distillation loss function with an uncertainty-aware weighting term to down-weight noisy or artifact-heavy training trajectories automatically.
The paper presents LiveTalk, a real-time system for interactive multimodal avatar video generation using a novel video diffusion distillation recipe. By refining multimodal data quality, ensuring full ODE convergence before distillation, and using an aggressive optimization schedule, the authors achieve 20x acceleration and sub-second latency while maintaining visual quality comparable to larger bidirectional models.
The core technique is an improved 'Self Forcing' on-policy distillation method for Distribution Matching Distillation (DMD). The system uses a causal, block-wise autoregressive video diffusion model conditioned on audio, text, and images, integrated with 'Anchor-Heavy Identity Sinks' (AHIS) in the attention mechanism to preserve speaker identity over long streaming sessions.