Google DeepMind reconstructs Pelé's "lost" 1959 goal using Veo, Gemini Omni, and Nano Banana Pro

Google DeepMind has digitally reconstructed a goal scored by Pelé on August 2, 1959 — the "Gol da Rua Javari" — that was never captured on film, using a combination of practical filmmaking and generative AI models. The reconstruction, published July 14, 2026, was built in collaboration with Pelé's family, football historians, sports journalists, and eyewitnesses, and is now on permanent display at the Pelé Museum in Santos, Brazil.

What's new

DeepMind's team pieced together the play from more than 3,600 historical images, period accounts, and eyewitness testimony, since no video or photograph of the goal itself is known to exist. To bring it to life, the team combined stunt-player choreography and 3D geometry/motion capture with three of Google's newest generative models:

• Veo, used with Performance Control to direct and generate the match footage
• Gemini Omni, used for multimodal editing and refinement across the sequence
• Nano Banana Pro, used for character replacement and environment restyling to match 1959-era Santos

As DeepMind put it: "Recreating this goal required a combination of practical filmmaking and our most advanced AI models: Veo, Gemini Omni and Nano Banana Pro." The result is a hybrid post-production pipeline — practical stunt footage shot first, then reworked frame-by-frame with AI to match players' likenesses, period-accurate kits, the Vila Belmiro stadium environment, and the ball's actual scoring trajectory as described by witnesses.

Context

The project is the latest in a string of DeepMind efforts applying its media-generation stack to real-world archival and cultural problems rather than pure benchmark demos. It follows DeepMind's other recent applied work — including AlphaEvolve rollouts for Google Cloud customers and a research partnership with A24 — and continues Google's push to position Veo and Gemini Omni as production-grade tools rather than novelty generators. Performance Control, the Veo feature used to direct the reconstructed match action, has been positioned by Google as a tool for choreographing specific physical movements rather than leaving motion entirely to a text prompt.

Why it matters

The Pelé reconstruction is a concrete test case for a use of generative video that's harder to dismiss than typical AI slop: filling a real historical gap using verified eyewitness and documentary evidence as ground truth, with named institutions (the Pelé Museum, Santos FC's history, historians) attached to the result. It's also a signal of where Google intends its video/multimodal stack to be judged — not just on leaderboard scores, but on whether the output can survive scrutiny from experts and family members and stand as a permanent museum exhibit. For a technology still fighting a credibility problem around synthetic video, that's a meaningfully higher bar than most AI demos attempt.