Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation

Video generation has recently emerged as a central task in the field of generative AI. However, the substantial computational cost inherent in video synthesis makes model distillation a critical technique for efficient deployment. Despite its significance, there is a scarcity of methods specifically designed for video diffusion models. Prevailing approaches often directly adapt image distillation techniques, which frequently lead to artifacts such as oversaturation, temporal inconsistency, and mode collapse. To address these challenges, we propose a novel distillation framework tailored specifically for video diffusion models. Its core innovations include: (1) an adaptive regression loss that dynamically adjusts spatial supervision weights to prevent artifacts arising from excessive distribution shifts; (2) a temporal regularization loss to counteract temporal collapse, promoting smooth and physically plausible sampling trajectories; and (3) an inference-time frame interpolation strategy that reduces sampling overhead while preserving perceptual quality. Extensive experiments and ablation studies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion realism. It consistently outperforms existing distillation baselines across multiple metrics.

Spectral Roll-off Points: Estimating Useful Information Under the Basis of Low-frequency Data Representations

Useful information is the basis for model decisions. Estimating useful information in feature maps promotes the understanding of the mechanisms of neural networks. Low frequency is a prerequisite for useful information in data representations, because downscaling operations reduce the communication bandwidth. This study proposes the use of spectral roll-off points (SROPs) to integrate the low-frequency condition when estimating useful information. The computation of an SROP is extended from a 1-D signal to a 2-D image by the required rotation invariance in image classification tasks. SROP statistics across feature maps are implemented for layer-wise useful information estimation. Sanity checks demonstrate that the variation of layer-wise SROP distributions among model input can be used to recognize useful components that support model decisions. Moreover, the variations of SROPs and accuracy, the ground truth of useful information of models, are synchronous when adopting sufficient training in various model structures. Therefore, SROP is an accurate and convenient estimation of useful information. It promotes the explainability of artificial intelligence with respect to frequency-domain knowledge.