Colorize Image (DDColor)

How it fits

Typical backends

Caveats

• Hard constraintThe DDColor weights are not bundled in the worker; a model bundle (a directory holding a config.json plus a single safetensors or checkpoint weights file, e.g. the piddnad/ddcolor_modelscope snapshot) must be bound to the model file slot. The model is loaded once at startup directly from that bound directory with no network access.
• AccuracyColor is generated rather than recovered, so the model invents plausible hues; the predicted colors are not guaranteed to match the true original colors of the scene.
• I/O contractThe network runs on a fixed square input internally but the predicted color is mapped back onto the source resolution, so the output keeps the input's dimensions and only the chrominance is changed. A colour input is returned in its original color space; a single-channel grayscale input is returned as a 3-channel color image, since the predicted color cannot be carried by a grayscale frame.
• FallbackAn empty image is passed through unchanged, and requesting a GPU with none present falls back to CPU where per-image latency rises sharply.
• State lifetimeThe bound weights and the input size are resolved once at startup, so switching models or resolution requires a redeploy.