bespoke AI – Films at this year’s Tribeca Film Festival offered a clear signal: generative AI can help—when it’s treated like a specialized tool inside a human-led production workflow, not a plug-and-play replacement for filmmaking.
For Hollywood, the pitch has been loud and persistent: generative AI would remake filmmaking. But when you sat through the experiments at this year’s Tribeca Film Festival. the most striking thing wasn’t what AI could magically deliver on its own. It was what kept breaking—and what filmmakers did to work around it.
A chunk of the festival lineup looked like a warning label for anyone expecting studios to crank out feature-ready work just by feeding prompts into “vanilla” models. Roar. an animated short produced by Illuminai Studios. read more like a disorienting montage of AI-generated clips than a cohesive piece of cinema. Asteria Film Co.’s ChikaBOOM!. aimed for the fast. punchy energy of fantasy about a magician in training. but it lacked the visual and sonic polish that would be needed to pull viewers in.
The roughness wasn’t hard to connect to the production approach. In Roar and ChikaBOOM!, the projects leaned heavily on AI-forward pipelines, and the results reflected the limits of today’s generative video systems—short bursts of footage that often don’t hold together visually.
Then came the counterpoint that felt more like a blueprint than a dream.
Google DeepMind’s Dear Upstairs Neighbors. written and directed by Pixar veteran Connie Qin He in collaboration with researchers from Google DeepMind. follows an exhausted young woman trying to go to bed. All Ada wants is a couple hours of peaceful rest before she has to wake up and get back to work. But as she begins dozing off. the cacophony of noise from her upstairs neighbors’ apartment jolts her awake. leaving her wondering what could possibly be happening in the middle of the night.
He didn’t just design the story; she helped define its look. To give the world a distinct style. she enlisted Pixar production designer Yingzong Xin. who painted concept art in Photoshop and on paper using acrylics. Those illustrations come with an expressionistic aesthetic that was central to bringing the story’s fantastical feel to life. But that painterly look created a practical problem for the video models: with many AI video generation systems. that kind of stylized output is difficult to reproduce consistently.
DeepMind’s engineers developed custom versions of Veo and Imagen built specifically to let the film’s artists fine-tune their outputs. Because the customized models were trained on Xin’s concept art. they could generate shots that adhered to He’s vision for the project. The text-to-video models also handled stylistic details—like the way sound is visualized when objects interact—yet the short’s creative team still had to use more traditional methods to make the scenes land as a cohesive story.
The production team created rough animations with Autodesk Maya, described as an industry standard for 3D rigging and VFX. Those roughs were then fed into Veo so the artists could produce more visually polished scenes. which could then be enhanced further with additional stylized assets generated with Veo and Imagen.
That blend—customized models on top of deliberate. human-made animation—was what made Dear Upstairs Neighbors feel like something different from the usual “AI video” experience. It didn’t give the impression of a generator replacing artistic decisions. The workflow relied on human-made art and nuanced creative choices that text-to-video generators aren’t capable of on their own.
The difference mattered. because Dear Upstairs Neighbors wouldn’t have been nearly as impressive using vanilla versions of Google’s models. The film itself served as a kind of commercial for Google’s technology. but it also made the limits obvious: the models worked for this specific short. and that specificity wasn’t subtle.
OpenAI’s contribution at Tribeca felt. in comparison. harder to square with the idea that AI video is ready for prime-time storytelling. Alice Gu’s semi-autobiographical drama Smoked used Sora to re-create the Palisades Fire. Youssef Michraf’s Mauvais Soleil featured a number of photorealistic scenes generated with OpenAI’s creative tools.
Even with those aims, the films still showed the familiar workarounds. In Smoked’s fiery wide shots. the effect looked a bit cartoony—yet it worked better in close-ups of a woman and her son trying to escape the blaze in their car. Those close-ups were filmed using a Volume-like setup. Mauvais Soleil took a different tack: many shots last only a few seconds. and the only speaking character is an unseen narrator. which made those limitations feel like intentional artistic choices rather than unresolved technical constraints.
OpenAI’s presence also carried an uncomfortable detail. The company had recently decided to shut Sora down entirely. That decision is what led to OpenAI’s feature-length film Critterz not being able to debut at this year’s Cannes Film Festival. The timing left a message many in the industry would understand without anyone spelling it out: video-focused bets can disappear fast.
At the same time, the festival didn’t suggest the end of film experiments with gen AI—just a narrower, more disciplined role for it. The projects at Tribeca suggested workflows only work when they’re guided by human artists with very clear creative visions.
For a smaller-scale example. writer/director Ash Koosha used Kling AI. Claude. Gemini. and Nano Banana to singlehandedly produce Dreams of Violets. The docudrama focuses on nationwide protests that have rocked Iran throughout the past year. and Koosha tells a fictionalized story about a group of people trapped in an alley as police stalk the streets brutalizing civilians. The project took Koosha just a few weeks to finish and reportedly cost only $2,000 in computing costs. It’s supported by a strong narrative, even if it doesn’t break new ground visually.
Taken together. the message at Tribeca wasn’t subtle—Hollywood’s heavyweights aren’t likely to sign their names to commercially viable projects by handing prompts to generative models and hoping for the best. That kind of “video slop” may keep existing. but the bigger shift seems to be toward a different partnership model: major AI firms working with studios to build bespoke models tailored to very specific workflows.
And even if that future arrives, it still looks like the human part comes first. The work. across the festival’s strongest examples. was guided by artists making nuanced creative calls—while the technology handled the parts it could do well: generating assets within carefully controlled parameters. under direction. and inside a pipeline that already knows what the story must become.
generative AI Hollywood Tribeca Film Festival Google DeepMind Veo Imagen OpenAI Sora filmmaking AI video models Kling AI Claude Gemini Nano Banana