QuantWM: Temporally Consistent 2-Bit KV Cache Quantization for Video World Models
QuantWM applies training-free 2-bit KV cache quantization to video world models while preserving temporal consistency.
Video world models store growing KV caches for long-range consistency, but existing 2-bit KV quantization that looks nearly lossless on VBench causes flickering and visual degradation on these models. The authors find Key quantization has smaller reconstruction error than Value yet larger output damage, because Key perturbations shift attention logits and which temporal-spatial tokens Queries select. QuantWM is training-free: quantization-sensitivity-aware clustering picks INT2-friendly Key centroids, and principal-subspace attention compensation corrects residual Key error along the dominant Query subspace. On LingBot-World-v2, HY-World 1.5, Matrix-Game-2, Longcat-Video, and Causal-Forcing it improves quality and consistency with up to 6.20× KV cache compression.
- Training-free 2-bit KV cache quantization for video world models.
- Key perturbations shift attention logits and selected spatiotemporal tokens.
- QSAC clustering and PSAC compensation stabilize attention during quantization.
- Up to 6.20× KV memory compression with limited extra overhead.
From the original article
1 min read at huggingface.co
Video world models achieve long-range temporal consistency by storing KV cache during generation, but the growing cache makes KV cache memory a major deployment bottleneck, which motivates low-bit quantization study for efficiency.
By Jiaqi Zhao, Xiaobin Hu, Bo Yin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.