Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models

2026-05-10

AI Summary

This work addresses the high computational complexity and memory overhead in autoregressive video diffusion models, which stem from redundant key-value (KV) cache accumulation across historical frames and hinder scalability. To mitigate this, the authors propose a hybrid KV cache compression strategy that leverages the functional specialization observed among attention heads. Specifically, attention heads are partitioned into static and dynamic categories, with structured pruning applied to static heads and a similarity-driven dynamic pruning mechanism employed for dynamic heads. This approach substantially compresses the KV cache while preserving generation quality. Evaluated on a single NVIDIA H200 GPU, the method achieves over 29 FPS inference speed with a 30% reduction in cache memory usage, yielding 1.50× and 2.82× speedups over the baseline at 480p and 1080p resolutions, respectively.

Abstract

Autoregressive (AR) video diffusion models adopt a streaming generation framework, enabling long-horizon video generation with real-time responsiveness, as exemplified by the Self Forcing training paradigm. However, existing AR video diffusion models still suffer from significant attention complexity and severe memory overhead due to the redundant key-value (KV) caches across historical frames, which limits scalability. In this paper, we tackle this challenge by introducing KV cache compression into autoregressive video diffusion. We observe that attention heads in mainstream AR diffusion models exhibit markedly distinct attention patterns and functional roles that remain stable across samples and denoising steps. Building on our empirical study of head-wise functional specialization, we divide the attention heads into two categories: static heads, which focus on transitions across autoregressive chunks and intra-frame fidelity, and dynamic heads, which govern inter-frame motion and consistency. We then propose Forcing-KV, a hybrid KV cache compression strategy that performs structured static pruning for static heads and dynamic pruning based on segment-wise similarity for dynamic heads. While maintaining output quality, our method achieves a generation speed of over 29 frames per second on a single NVIDIA H200 GPU along with 30% cache memory reduction, delivering up to 1.35x and 1.50x speedups on LongLive and Self Forcing at 480P resolution, and further scaling to 2.82x speedup at 1080P resolution. Code and demo videos are provided at https://zju-jiyicheng.github.io/Forcing-KV-Page.

Problem

• autoregressive video diffusion
• KV cache compression
• attention complexity
• memory overhead
• scalability

Innovation

• KV cache compression
• autoregressive video diffusion
• attention head specialization
• hybrid pruning
• efficient generation