The Flow has not summarised this story yet — read it at arXiv Computer Vision.
arXiv:2610.20111v1 Announce Type: new Abstract: Distilling pretrained video diffusion models into few-step autoregressive (AR) generators typically combines student initialization with distribution m...
arXiv:2609.36995v1 Announce Type: cross Abstract: Few-step streaming audio--video generation requires both causal modeling and step distillation, yet standard training recipes face two context-relate...
arXiv:2610.26025v2 Announce Type: replace-cross Abstract: Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two...
arXiv:2608. 01263v1 Announce Type: new Abstract: On-policy distillation (OPD) samples trajectories from the current student policy and minimizes token-level divergence between student and teacher next-token distributions at prefixes along those trajectories.
TAM (Task‑Aware Memory Distillation) is a framework that stores a frozen teacher’s knowledge in a bounded, retrievable memory, encoding latent features, forecast changes, or flow residuals. During training, the student matches the teacher’s similarity distribution over shared references or regresses observation‑conditioned residual prototypes, complementing supervised prediction and conventional distillation. Experiments on video prediction, weather forecasting, and traffic flow prediction show that TAM improves SSIM on all six video datasets and reduces MSE on five, achieving up to 1.93% MSE reduction on WeatherBench without affecting student inference cost.
We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout.