How do I colorize an old black and white photo?
Open the photo in a colorization tool and let a trained model predict the colours. The original ones are gone, so every result is a plausible reconstruction rather than a record. Repair damage first, colorize second, enlarge last. QuietPix runs the model on your device, so the photo never leaves your computer.
- Digitise the print properly. A flatbed scan at 300–600 DPI gives the model the most to work with. Photographing the print works too: lay it flat, use even indirect light, and keep the flash off — glare erases the tonal detail colour is inferred from.
- Repair the damage before adding colour. Scratches, dust and creases are read as image content, so a torn sky can come back tinted along the tear. Run the photo restorer first; colorize the repaired copy.
- Colorize. Drop the photo into the colorizer. The model downloads once on first use and is cached afterwards, so the second photo starts immediately. Everything after that runs on your own machine.
- Enlarge last, then save. Upscale after colorizing rather than before. The upscaler works from a finished image, and enlarging first only makes the colour step read a bigger version of the same detail.
A colour version of the photo in a few seconds per image after the one-time model download, at the original pixel dimensions. Keep the black and white file: colorization is an interpretation, and you may want to redo it.
Sepia and yellowed prints need no preparation
A common instruction is to convert an aged photo to grayscale before colorizing. That step is unnecessary here. The image is decomposed into Lab and only the L channel — its brightness — is handed to the model, so a sepia cast, a yellowed emulsion or a faded blue tint is discarded on the way in. Feeding the tool an untouched scan of an old print is the correct move.
What the model is reliable about, and what it invents
Skin, sky, foliage and wood come back convincingly, because those things have a narrow range of plausible colours and the model has seen them constantly. Clothing, painted walls, cars and flowers do not — a dress that reads mid-grey could have been almost anything, and the model settles on the most statistically ordinary option. Treat those areas as a guess rather than a discovery, especially before telling a family member what colour their coat was.
Why the result stays sharp at 512 pixels of model
The network itself works on a fixed 512×512 input, which sounds like it would soften a large scan. It does not, because only the colour is taken from the model output. The finished image recombines your original full-resolution brightness channel with the predicted colour, upsampled underneath it. Chroma is low-frequency information — the eye tolerates it being soft — while every bit of grain and edge detail in the scan survives untouched.
Tools for this job
Frequently asked questions
Are the colours historically accurate?
No, and no tool can make them so. The colour information was never recorded by black and white film; a model predicts what was most likely. It is a plausible reconstruction, useful for seeing a face as a face rather than for documenting what someone wore.
Does my photo get uploaded anywhere?
No. The model file is fetched once from a public model host, then the colorization itself runs in your browser on your own hardware. The photo is never transmitted, which also means the tool keeps working with the network off after that first download.
Can I colorize a drawing, a document or a painting?
You can try, but expect weak results. The model was trained on photographs, so line art, scanned text and engravings fall outside what it knows and usually come back barely tinted.
The result looks washed out. What went wrong?
Usually a low-contrast source. If the scan is flat and grey, the brightness channel carries little for the model to key on. Correct the contrast of the black and white file first, then colorize.