Photo Colorization Lab
The trained SIGGRAPH 2017 model receives photo lightness (L*) together with your color hints and mask, and predicts color components (a*b*). The proposal preserves the original lightness and keeps hint centers at their requested color.
Key features
- Verify a pinned SHA-256 for the self-hosted trained model before browser WASM inference
- Feed photo L* plus user a*b* hints and a hint mask as separate model inputs
- Add and remove up to 20 color/radius hints by canvas click or numeric coordinates
- Export a PNG with source lightness, a transparent hint map and settings JSON
- Cancel loading or inference by terminating the worker and discarding stale results
How to use
- Choose an opaque 64×64 to 1600×1200 photo or open the built-in sample.
- Pick a hint color and radius, then click the preview or enter X/Y coordinates.
- Run the model and wait for local download and inference.
- Inspect the proposal, adjust hints and rerun as needed, then save PNG, hint map and JSON.
Use cases
- Try several possible clothing colors in a monochrome portrait
- Steer sky and foliage colors in an old landscape photo
- Compare the grayscale source with a learned color proposal before manual restoration
- Keep exact hint positions and model hash in a JSON record
Frequently asked questions
Are the proposed colors historically accurate?
No. The trained model proposes plausible colors and cannot recover the actual original colors. Do not treat the proposal as historical evidence.
Is this a sepia or tint filter?
No. The trained neural network takes source lightness plus color-hint values and mask, and predicts image-specific color components. Its output changes when hints change.
What happens to large photos?
The model infers at 256×256, then color components are interpolated to the source size while source lightness is retained. Fine edges may have color bleed.
Is my photo uploaded?
No. Photo decoding and inference run in your browser. Only fixed model and WASM code files are downloaded from this site.
Privacy
Photos and hints are processed in a browser worker. Your browser downloads a roughly 34 MiB model and 13 MiB local WASM runtime from this site, but your photo is not uploaded.
Comments & questions