Google DeepMind has some of the best video models going around, and it’s also collaborating with the industry to tweak them for real-world use cases.
The research lab and A24 — the production company behind films like Midsommar, Everything Everywhere All at Once, and Hereditary — have announced what they’re calling a first-of-its-kind research partnership. The collaboration is structured as a deep R&D relationship spanning multiple projects over time, with Google also making a direct investment in A24 as part of the deal.
The stated goal is to embed Google DeepMind’s research capabilities directly into A24’s creative process, so filmmakers can help shape tools before those tools are finalized — rather than receiving finished products and being expected to adapt. A24’s directors and producers would work side-by-side with DeepMind researchers to test, iterate, and build, with the creative feedback flowing back into the research process.
That framing matters. AI companies have tended to build video generation tools and then go looking for use cases. This partnership inverts that structure. Veo 3.1, DeepMind’s current video generation model, is already considered one of the most capable in the space — capable of generating synchronized audio alongside high-fidelity video. But raw capability and professional filmmaking utility are different things, and A24’s involvement is an attempt to close that gap.
A24 is a deliberate choice of partner here. The studio has a reputation for backing unconventional projects and giving directors unusual levels of creative control, which makes it a reasonable testing ground for technology that doesn’t yet have a defined role in production pipelines. The partnership doesn’t lock down specific deliverables or technical milestones — both sides have been explicit that goals will evolve as the collaboration develops.
For Google DeepMind, the value is the feedback. Building video models in a lab environment is a fundamentally different exercise from deploying them in a context where professional filmmakers are relying on them to realize a specific creative vision. The gap between what a model can do on a benchmark and what it can do on a film set has never been properly tested at this level.
Whether the partnership produces tools that actually change how films get made remains to be seen. But the structure — researchers embedded with filmmakers over a sustained period, rather than a one-off demo or a licensing deal — is a more serious approach to the problem than most of what’s come before.