VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
2026-05-28 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial Intelligence
AI summaryⓘ
The authors studied a new way to handle memory in video diffusion models called Multi-Head Latent Attention (MLA). Instead of keeping separate keys and values for each attention head, they use a shared compact representation that saves a lot of memory. They found that even though video attention data doesn’t fit certain mathematical assumptions from language models, their approach still works well and keeps video quality. Their method, VideoMLA, performs as well or better than current methods on benchmarks while using less memory and running faster.
video diffusioncausal attentionkey-value cachemulti-head attentionlow-rank approximation3D-RoPE positional encodingmemory efficiencystreaming videoeffective rankB200 GPU
Authors
Hidir Yesiltepe, Jiazhen Hu, Tuna Han Salih Meral, Adil Kaan Akan, Kaan Oktay, Hoda Eldardiry, Pinar Yanardag
Abstract
Long-rollout causal video diffusion has converged on a fixed-size sliding-window KV cache, with recent progress innovating within this layout by changing which tokens occupy the window or how their positions are encoded. The per-head KV layout itself, a dominant contributor to streaming memory and latency, has been mostly left unchanged. In this paper, we present the first study of Multi-Head Latent Attention (MLA) in video diffusion. VideoMLA replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7% at every cached layer. We further investigate why MLA succeeds in video diffusion even though the spectral assumption often used to motivate it in language models does not hold: pretrained video attention is not low-rank, with 99%-energy effective rank far above any practical latent dimension. VideoMLA retains quality at compression ratios where direct spectral approximation would predict large reconstruction error. We show that the MLA bottleneck, rather than the pretrained spectrum, determines the effective rank: both spectral and random initialization occupy nearly the full rank budget from initialization, and training preserves this budget while adapting within it. On VBench, VideoMLA matches short-horizon streaming video diffusion baselines, achieves the best overall score at long horizons among evaluated methods, and improves throughput by 1.23x on a single B200.