QuantWM addresses temporal flickering in video world models caused by standard 2-bit KV cache quantization. It identifies that Key perturbations in video models shift temporal-spatial token selection, leading to visual degradation, and proposes a more stable quantization approach.
HOW THIS AFFECTS YOU
●
builderUse this to achieve 2-bit KV cache efficiency in video models without sacrificing temporal consistency.
●
researcherThis provides insight into why Key quantization causes larger output degradation in video-centric attention mechanisms.