Video that keeps the cast looking like themselves

On Sep 24 Google Research published work on an AI video co-director, a multi-agent orchestration layer on Gemini and Veo for coherent long-form, multi-shot video. Companion frameworks CANVAS, A²RD, and VQQA handle visual memory, minutes-long generation, and closed-loop prompt fixes. The stack inherits SynthID. This is research headed to academic venues, not a consumer button yet.

What's interesting: Short clips already look good. The hard part is the middle of the film, when identity and rooms start to melt. Google is attacking that with world-state tracking and a judge loop instead of hoping a longer diffusion window holds.

For people who make brand films, product stories, or narrative spots, the craft signal is not that AI made a movie. It is that continuity labor might finally get a co-pilot. Someone still has to own the story, the taste, and the cut. The tool is useful when it stops the characters from mutating so the human can keep directing.

Primary: https://research.google/blog/coherent-long-form-video-generation/