AdaState: Self-Evolving Anchors for Streaming Video Generation

AdaState introduces a self-updating latent anchor for streaming video diffusion models.

It replaces the fixed first-frame reference with a hidden state in the KV cache that evolves over chunks.

The state is denoised and refreshed using both past state and current content, reducing attention lock on the first frame.

This improves temporal continuity and helps video generation adapt as scenes change.