Video generation is hitting a wall. As models attempt to create longer, more coherent clips, the memory demands of tracking context explode, forcing a trade-off between duration and speed. Researchers have now cracked a key bottleneck, enabling minute-scale video diffusion without the usual lag.

The team introduces VideoMLA, a method that drastically cuts the memory footprint of video generation. Standard models store separate data for every attention head, which consumes massive amounts of RAM. VideoMLA replaces this with a shared low-rank content layer and a simplified positional key. This clever compression slashes KV cache memory usage by 92.7 percent, allowing the model to process far more information per token.

The surprise lies in why this works. Theory suggests such heavy compression should destroy quality, as video attention data is not naturally low-rank. Yet, VideoMLA maintains high fidelity. The researchers found that the bottleneck itself, not the original data structure, dictates performance. Training adapts the model to fit within this tight budget, preserving visual quality even at compression ratios that should theoretically fail.

The results are tangible. VideoMLA matches existing short-video baselines while delivering the best overall scores for long-horizon generation. It also boosts processing speed by 23 percent on high-end hardware. This breakthrough proves that efficient memory management is just as critical as model architecture for the future of generative video.