Colorize Black-and-White Photos

In your browser. The colorization model (~55 MB Standard, ~110 MB High) is downloaded once via ONNX Runtime Web and cached for reuse, then runs on WebGPU where available and falls back to WebAssembly on older browsers.

It happens the first time you colourize, not when the page opens, and it is about 55 MB for Standard or 110 MB for High. That is a one-off: the file is cached afterwards, so a second photo starts painting immediately. Standard is the sensible default — the difference shows mainly on large scans with a lot of fabric and foliage, where High resolves finer boundaries between colors. On a machine with WebGPU the work runs on the graphics card and finishes in seconds; without it the same model still runs, just more slowly.

It can, and it is the most common reason a result comes out muddy. The network works from brightness and paints color onto it, so a scan that has gone yellow or sepia is feeding it a color cast to interpret as information — and it obliges, usually by making everything warm. Desaturating the scan to true gray before colourizing removes the problem at the source. If the faces still come out sallow, the skin-tone slider is there for exactly that, and a reference photo of at least 64×64 can hand the model a palette to borrow from.

Yes — switch the content type to Cartoon/Anime, which uses a model trained on line-art instead of photographs. It works best in a WebGPU browser; on devices without WebGPU it falls back to the photo model.

AI colorization is an educated guess and skin or saturated objects can be wrong. Use the skin-tone slider, drop a reference photo, or use "Fix wrong colors" to recolour or desaturate a specific hue across the whole image.