Jul, 134 min to read
Platform
News · Google's Wizard of Oz restoration at Sphere: a fine-tuning problem disguised as a spectacle
Apr, 84 min to read
PlatformGoogle's Wizard of Oz restoration at Sphere: a fine-tuning problem disguised as a spectacle
How Google DeepMind and Google Cloud are stretching three grainy 1939 film negatives across a 160,000-square-foot LED screen using Imagen, Veo and Gemini
Three film negatives, 160,000 square feet
The technical starting point is unusually specific. Buzz Hays, Google Cloud's global lead for entertainment industry solutions, describes the source as "the original four-by-three image on a 35mm piece of celluloid — it's actually three separate, grainy film negatives; that's how they shot Technicolor."
That source has to land on Sphere's 16K LED screen — described as the highest-resolution screen in the world — across 160,000 square feet of surface, wrapping a 17,600-seat venue. The mismatch between a 1939 4:3 negative and a spherical wraparound canvas is the entire problem the project exists to solve. It debuts August 28.
The framing here matters: this is not upscaling for a bigger TV. A conventional theatrical frame constantly cuts between characters, removing them from view. On Sphere, everything in a scene has to coexist at once, which means the missing elements have to be reconstructed, not just enlarged.
Three tuned models doing three different jobs
Google breaks the pipeline into three distinct operations, each handled by task-specific versions of Veo, Imagen and Gemini. Super resolution converts the tiny celluloid frames into ultra-high-definition imagery. Outpainting expands scenes to fill space and to fill the gaps left by camera cuts and framing. Performance generation composites the original performances into those expanded environments.
The interesting platform detail is that these aren't off-the-shelf model calls. Each stage uses a version "specially tuned for the task." The announcement is effectively describing a media-generation stack where distinct models are specialized per-function rather than a single model asked to do everything through prompting.
DeepMind researcher Dr. Steven Hickson's account of the timeline is candid about how unfinished this was at the start: "We'd find something we can't do, we think it's impossible, and then a month later we're like, actually, maybe we can do that."
The archive as training data
The most transferable idea in the piece is that the biggest quality gains came not from a better model but from more source material. The team scoured archives for the shooting script, production illustrations, photographs, set plans and scores, then fine-tuned Veo and Gemini on those materials.
Notably, the fine-tuning corpus includes production metadata like camera focal lengths for specific scenes — the kind of detail that constrains a generated frame to match how the original was actually shot. Google reports concrete outputs: "Dorothy's freckles snap into focus and Toto can scamper more seamlessly through more scenes."
This inverts the usual generative-media story. The lever wasn't a smarter architecture; it was assembling a dense, domain-specific dataset well beyond the 102-minute film itself, so the model learned this particular production rather than film in general.
A hard constraint: no new material
The project imposes a rule that shapes every technical choice: no new dialogue was written and no new music was recorded. Producer Jane Rosenthal says the team considered other approaches before concluding "we really needed to do it with AI," and Hays notes every change was made with Warner Bros. to preserve continuity with the original.
For an applied team, this is the useful signal. The value wasn't generation for its own sake — it was generation bounded by an existing, licensed, canonical work, where the model's job is to extend rather than invent. That constraint is what made fine-tuning on archival production materials essential, and it's what separates this from a system that could confabulate freely.
The specific implication: Google is positioning Imagen, Veo and Gemini not as text-to-video toys but as a restoration-and-extension pipeline for rights-holders who need output that stays faithful to a fixed source. The differentiator on display is task-specialized tuning plus disciplined data curation — not the raw generative capability alone.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.