How did Google DeepMind recreate a 70-year love story that was never filmed?

The couple and the memory that had no photo

Burt and Ethelle Shatz have been married for over 70 years. Burt now lives with cognitive decline. One memory refuses to leave them both — the day they met at a student co-op in Cleveland. No camera captured it. No film reel exists. The memory lived only in two aging minds.

Director Liz Garbus wanted to make that memory visible. She brought in Google DeepMind and Primordial Soup. The project became a collaboration between engineers and filmmakers. Michael Chang served as technical lead. His own grandfather had suffered a stroke and memory loss before he passed. The last time Chang visited, his grandfather was convinced he was still in college and graduating. That bittersweet moment never left him. When the project began, Chang asked his father for old family photos. He tested image restoration on photographs from when his parents met and used video models to animate them. Watching his parents move as twenty-somethings drove home how these tools could keep what matters most from slipping away.

Two technical layers built the scene

The reconstruction relied on a two-part technical approach. First, generative models restored black-and-white photos of Burt and Ethelle from their youth. The restored photographs anchored every later frame. Second, performance capture models mapped present-day micro-mannerisms onto those younger likenesses. The specific tilt of Burt's head. A brief hesitation in his speech pattern. The subtle crinkle around his eyes. Engineers fed these traits into the model. The output merged past appearance with present behavior. The result felt authentic to the couple.

Ethelle guided every frame

Ethelle sat alongside the engineering team. She corrected the curve of a staircase. She adjusted the shape of a shoe heel. Her voice directed the AI like a painter directs a brush. Darren Aronofsky compared the technology to a paintbrush. It does nothing until human hands guide it. The film demonstrates that principle frame by frame.

Liz Garbus had seen fMRI scans light up when patients heard familiar voices. She had also seen scans respond to images of loved ones. Years later Darren Aronofsky encountered footage of a former ballerina with Alzheimer's. The dancer instinctively performed Swan Lake choreography from her wheelchair. Those observations led the creative team to reminiscence therapy. That clinical practice uses sensory cues to stimulate memories. Songs serve as one cue. Family stories serve as another. Old photographs serve as a third.

From documentary to Gemini app

Anyone can try a version of this restoration. Open the Gemini app. Upload an image of an old photograph. Ask: "Can you restore and colorize this photo? Preserve the appearance, expression, and pose of the people." The app uses similar generative models. It does not animate video. It works on still images only. The film itself is available to watch online. A link appears at the end of the Google blog post.

Where the explanation stops

The documentary does not disclose the compute budget. It does not name the specific model versions. It does not measure accuracy against ground truth. No clinical trial validates the therapeutic effect. The couple's reaction appears on screen but their private assessment remains private. The technology works for this one story. Whether it scales to other families with different photo archives is an open question.