Imagen earns its place on style consistency
Technology
An image model inside a game is not making art. It is making a game object that has to sit inside a scene something else already generated, in the same style, without looking pasted on top of it. Consistency matters more than fidelity here.
Living Canvas, the playable web game we built with Google for Google I/O 2025, runs on that constraint. A player draws something, Imagen renders it, and the object joins a scene already running in the style the player picked.
Why
What Imagen actually does in this build
- The input is a drawing, not a prompt
A player sketches a shape and expects that shape back. Whatever Imagen returns has to be recognizably the thing they drew, in the place they drew it, or the puzzle stops making sense. That is a much tighter brief than an open image request.
- One style has to carry across everything
Living Canvas scenes run in styles like realistic, cartoon, and pixelated, and the style chosen for the background carries through to every object drawn afterward. A car and a rain cloud generated minutes apart still have to belong to the same picture. That is the hard part.
- Consistency is carried, not requested
An image API has no memory of the last call, so style is state the product carries rather than a setting the model holds. The question on every result is not whether it is a good image. It is whether it belongs in this one.
