“Dear Upstairs Neighbors” showcases Veo’s video-to-video AI workflows
PUBLISHED: Mon, Jan 26, 2026, 6:31 PM UTC | UPDATED: Mon, Jan 26, 2026, 6:44 PM UTC
- ■
Google DeepMind premiered “Dear Upstairs Neighbors” at Sundance, showcasing AI-assisted animation workflows developed with Pixar veteran Connie He
- ■
The team fine-tuned Veo and Imagen models on custom artwork, teaching AI to maintain consistent character designs and expressionist visual styles across shots
- ■
Novel video-to-video workflows let animators guide AI with rough animation instead of text prompts, maintaining precise control over timing and performance
- ■
Veo’s 4K upscaling launches in Google AI Studio and Vertex AI later this month, bringing professional filmmaking tools to wider audiences
Google DeepMind just premiered something unprecedented at the Sundance Film Festival – an animated short film created through a radical collaboration between traditional animators and AI researchers. “Dear Upstairs Neighbors,” screening at Sundance’s Story Forum today, represents a new frontier in creative AI tooling, where fine-tuned video models work alongside hand-crafted animation rather than replacing it. The project reveals how Google’s Veo AI can learn unique artistic styles and respond to visual direction, not just text prompts.
Google DeepMind just crashed Sundance with a wildly ambitious experiment. The animated short “Dear Upstairs Neighbors” premiered today at the festival’s Story Forum, but it’s not the charming story of a sleep-deprived woman battling noisy neighbors that has people talking – it’s how the damn thing got made.
Director Connie He, a Pixar alum, teamed up with Google’s AI researchers to create something that shouldn’t quite exist yet: a hand-crafted animated film where AI does the heavy lifting on the most technically brutal parts, while human artists maintain frame-by-frame creative control. The result challenges basically everything we thought we knew about AI’s role in creative work.
“The expressionistic visual styles are central to the storytelling – and extremely difficult to achieve in traditional animation,” supervising animator Cassidy Curtis explained in Google’s announcement. The team discovered early on that their vision was so specific, their styles so unique, that existing AI tools couldn’t cut it. So they built new ones.
Here’s where it gets interesting. The production team fine-tuned custom versions of Veo and Imagen – Google’s video and image generation models – by feeding them artwork from production designer Yingzong Xin. But this wasn’t simple style transfer. The AI learned deep artistic concepts like two-point perspective and how to maintain character silhouettes that follow 2D animation rules even as forms rotate in 3D space.
“What Veo learned from our concept art surprised us: not just superficial details like color and texture, but deep artistic concepts,” the team noted. When they showed the AI paintings of Ada, the protagonist, the model grasped that her characteristic hair poof must always appear in silhouette, never obscuring her face – a strictly two-dimensional rule that can’t exist in physical 3D space.
But fine-tuning alone couldn’t solve the control problem. Text prompts – the standard way to direct AI – proved useless for conveying comedic timing, the rhythm of sleepy fingers typing, or the exact framing of a camera reveal. So the team developed video-to-video workflows that let animators communicate the way they always have: by drawing, painting, or blocking out rough animation.
Animator Ben Knight created rough 3D animation in Maya for key scenes. Researcher Andy Coenen then fed that animation through fine-tuned Veo models, transforming gray mannequins into fully stylized, painterly characters that followed the exact motion and timing Knight intended. Other animators worked in TV Paint for 2D animation or acted out scenes themselves.
“We developed novel video-to-video workflows, which allowed our animators to convey their intentions visually,” Curtis explained. The models transformed rough animation into fully stylized videos “with an adjustable balance between tight control and creative improvisation.”
The workflow wasn’t magic. Early attempts at text-to-video generation with the fine-tuned Veo model produced scenes that looked like Ada but moved in “random, uncontrolled, and often bizarre” ways. The team’s blooper reel shows Ada eating spaghetti in ways that violate both physics and basic decorum.
Even with video-to-video control, no final shot emerged from a single generation. The team held daily “dailies” reviews just like traditional film productions, critiquing each shot through multiple rounds of feedback. They built localized refinement tools that let them edit specific regions of a video – adding an extra tuft of hair to improve Ada’s silhouette, for instance – without regenerating the entire shot.
For the film’s most intense expressionist sequences, the AI proved unexpectedly adept. Animator Steven Chao created dynamic low-poly effects in Maya. Researcher Ellen Jiang and director He then used fine-tuned Veo and Imagen models to transform those geometric shapes into lush splatters of neon paint. “The staccato rhythm of the changing paint texture adds to the intensity of the action,” the team noted.
Finally, they used Veo’s upscaling capability to bring shots to 4K resolution for theatrical presentation. The researchers tuned the upscaling model to add detail while preserving every nuance of the hand-painted aesthetic – no small feat when you’re trying to maintain the rough energy of expressionist brushwork.
The Veo 4K upscaling model launches in Google AI Studio and Vertex AI later this month, Google confirmed, making these professional filmmaking tools available beyond the research team.
What makes this project significant isn’t just the technical achievement. It’s the workflow philosophy. Instead of replacing animators with AI, Google DeepMind embedded researchers directly into the production process as technical artists. Animators got direct access to experimental research tools and helped shape their development through hands-on critique.
“Our artists found new creative powers through direct access to experimental research, and used their craft and perspective to help shape its development,” the team wrote. Meanwhile, researchers “gained hands-on experience as technical artists, rapidly prototyping solutions to break through artistic and technological barriers.”
This collaborative model stands in stark contrast to the AI industry’s usual approach of building tools in isolation, then releasing them to creative professionals with a shrug. The “Dear Upstairs Neighbors” team went through the painful, iterative process of actual production to understand what artists actually need – which turns out to be fine-grained control, not one-click magic.
The film arrives as Hollywood grapples with AI’s role in production. The 2023 WGA and SAG-AFTRA strikes put guardrails around AI use, but questions remain about how these tools will reshape creative workflows. Google’s approach – positioning AI as an augmentation tool that requires human direction at every step – offers one possible answer.
Whether “Dear Upstairs Neighbors” works as entertainment remains to be seen. But as a proof of concept for human-AI creative collaboration, it’s already accomplished something remarkable: showing that AI can handle the technically brutal parts of animation while artists maintain the creative vision that makes a film worth watching.
“Dear Upstairs Neighbors” matters less as a film than as a blueprint for what AI-assisted creative production could look like when done right. By embedding researchers into the production process and building tools around artists’ actual workflows rather than replacing them, Google DeepMind demonstrated that AI doesn’t have to be a black box that either does everything or nothing. The team’s video-to-video approach, fine-tuning capabilities, and localized refinement tools point toward a future where AI handles technical complexity while artists maintain shot-by-shot creative control. As these tools roll out to Google AI Studio and Vertex AI this month, the real test begins: whether this collaborative model can scale beyond a single experimental short to reshape how the industry approaches animation production.
Leave a Reply