Google and Pixar alums showcase AI-crafted short at Sundance

Ask any traditional animator about generative video, and you will usually get an eye-roll. Sure, the tech looks magical in Twitter clips, but on an actual production line, it is notoriously unusable: characters morph between frames, comedic timing gets lost, and text prompts cannot tell a character exactly when to blink.

Now, Google DeepMind is bringing an answer straight to the Sundance Film Festival. In collaboration with Pixar alum Connie He, the team just previewed "Dear Upstairs Neighbors," an animated short tracking a sleep-deprived woman battling the chaotic racket coming from the apartment above.

More than an art project, it serves as a live blueprint for how professional animators can tame generative video without sacrificing creative control.

Ditch the text box, show the motion

From early production, director Connie He and award-winning production designer Yingzong Xin realized that standard prompting could never capture the nuances of Ada's twitchy facial comedy or the rhythm of her fingers on a keyboard.

Instead of forcing artists to become "prompt engineers," DeepMind built a video-to-video pipeline. Animators sketched rough animations or acted out motions in their regular tools. DeepMind's models then transformed that rough blocking into fully rendered, painterly shots that adhered directly to the input motion, with a slider to balance strict control against AI improvisation.

To lock down visual consistency, researchers allowed the crew to fine-tune custom versions of Veo and Imagen on just a handful of concept paintings, ensuring the expressionistic visual style did not drift mid-scene.

No single-click miracles: bringing back the dailies

DeepMind is quick to admit that not a single final frame came out of a single prompt. The crew ran traditional daily review sessions, critiquing frames and sending shots back for notes.

Because hitting "regenerate" on an AI video usually scraps everything good about a take, researchers engineered localized refinement tools. Animators could highlight a specific region of a shot and tweak an arm swing or lighting quirk without re-rolling the whole sequence.

To prep the project for the theater, DeepMind leaned on Veo's 4K upscaler to enrich canvas textures without smoothing out stylistic brushwork. Google confirmed that this 4K upscaling tool is currently live in Flow and is slated to hit Google AI Studio and Vertex AI later this month.

Why it matters

For independent animators and visual storytellers, this pipeline offers a genuine sigh of relief. DeepMind did not treat AI as a slot machine that replaces labor; it turned it into a sophisticated render pass guided by rough hand-drawn drafts. If you know how to block out movement and stage a scene, these tools mean a skeleton crew could realistically produce short-film-grade visual fidelity without traditional multi-million-dollar rendering overhead.

For the broader AI race, this looks like a calculated strike against rivals like OpenAI's Sora and Runway. Tech demos on social media are not winning over Hollywood anymore. By embedding itself with Pixar talent at Sundance and preparing to roll out battle-tested 4K workflows onto Vertex AI, Google is positioning Veo as the mature enterprise infrastructure for professional media pipelines.

In our view, the smartest choice here was recognizing the limits of text. Prompting is dead for pro-grade filmmaking; visual inputs and localized keyframing are the only viable path forward. The real challenge now is productization: turning a white-glove research collaboration into intuitive, off-the-shelf software that any indie studio can run without a DeepMind engineer on speed dial.