Direct Causal-Mode Distribution Matching for Autoregressive Video Distillation
Apply a spatiotemporally weighted exponential decay mask on forward-KL loss to handle causal generation asymmetry without degrading structural quality.
Source paper
Kai Zou, Dian Zheng, Hongbo Liu, Tiankai Hang, Bin Liu, Nenghai Yu
Generated by AI. The paper’s authors did not write these.
Apply a spatiotemporally weighted exponential decay mask on forward-KL loss to handle causal generation asymmetry without degrading structural quality.
Integrate continuous memory compression into hierarchical video denoising to maintain an infinite-horizon context alongside a dense recent sliding window.
Replace hardcoded early-step KL regularization with a frequency-guided forward-KL loss that applies soft, step-dependent low-pass filtered targets across all denoising steps.
This paper proposes HiAR, an autoregressive framework for long video generation that mitigates distribution drift by utilizing hierarchical denoising. By conditioning generated blocks on context at matched noise levels instead of fully clean frames, it effectively reduces error propagation while enabling pipelined parallel inference.
The core methodology relies on a hierarchical causal generation schedule where all blocks are generated across denoising steps simultaneously. To prevent motion collapse during training, this is paired with a self-rollout distillation process utilizing a Distribution Matching Distillation (DMD) loss and a bidirectional forward-KL regularizer.