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

In three linesAutoregressive video diffusion models use quantized KV caches to reduce memory, but quantization creates an attention bias (Jensen bias) that degrades quality. Authors propose a per-attention-score correction computed from quantization step sizes, recovering quality lost with INT2 quantization while using 50% less memory than INT4.

Summary generated by Claude — human-verified