Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion

A recent study highlights the challenges faced by chunk-wise autoregressive video diffusion models, particularly regarding the memory bottleneck created by the KV cache as video lengths increase. The research identifies a systematic bias in attention weights, termed Jensen bias, which leads to a degradation in video quality due to quantization noise inflating the contribution of cached keys.

This development is significant as it proposes a per-attention-score correction method that mitigates the bias without incurring additional memory costs, potentially enhancing the performance of video generation technologies.

The findings underscore a broader trend in artificial intelligence where optimizing memory usage and computational efficiency is critical, particularly in autoregressive models. The introduction of frameworks like MotionCache further emphasizes the industry's focus on improving cache reuse strategies to address similar computational challenges in video generation.