When to Lock Attention: Training-Free KV Control in Video Diffusion

A team of researchers has introduced KV-Lock, a novel solution to a persistent problem in AI video generation: the trade-off between foreground quality and background consistency. When using diffusion models to edit videos—like changing a person's clothing or adding an object—injecting full scene information often creates artifacts in the background, while rigidly locking the background stifles creative changes to the foreground. KV-Lock tackles this by being a training-free, plug-and-play module that can be added to any pre-trained Diffusion Transformer (DiT) model.

The core innovation is its dynamic control system. KV-Lock monitors the AI's 'hallucination metric,' essentially the variance in its denoising predictions, which signals when the model is likely to generate inconsistent or low-fidelity content. When high risk is detected, the framework automatically increases the 'fusion ratio' of cached background information (Key-Values) and simultaneously amplifies the conditional guidance for the foreground subject. This dual-action approach mitigates background artifacts while empowering the model to generate higher-quality foreground edits.

Extensive experiments validate that KV-Lock outperforms existing methods across various video editing tasks. The framework provides a practical, efficient upgrade path for existing video generation systems like Stable Video Diffusion or Sora-style models, enabling more reliable and professional-grade edits without the computational cost of full model retraining.