Colorize black and white photos in your browser: drop in an old family photo, a scanned print or a sepia picture, and an AI colourisation model paints in plausible colours for skin, sky, grass, clothing and buildings in a few seconds. Drag the slider to compare before and after, turn the colour strength up or down until it looks right, and download the result at the photo's full resolution as PNG or JPG. There is no account, watermark or daily limit.
The model runs on your own device, so a photo of your grandparents never leaves your computer. Two honest limits: first, the colours are a learned guess, not the real colours of the scene. A grey coat could have been navy or brown, and the model cannot know; do not present the result as a historical record. Second, this tool only adds colour. It does not remove scratches, dust, creases or tears, and it does not sharpen a blurry photo. For a small or soft scan, run the AI Image Upscaler first, then colorize.
How to use
- Drop a black-and-white or sepia photo on the box, or click to choose one (JPG, PNG, WebP and other common formats, up to 12 megapixels).
- Press Colorize photo. The first time, the page shows what it needs to download (the 16 MB model and the 24 MB AI engine, less if another of our AI tools already loaded the engine) and starts when you agree; after that it is cached and starts immediately.
- Drag the before/after slider over the photo to compare the original with the colorized version.
- Adjust Colour strength: lower for a soft, hand-tinted look, higher if the colours look washed out. The change is applied instantly.
- Choose PNG (lossless) or JPG (smaller) and download. The file has the same pixel size as your original.
Frequently asked questions
How does AI photo colorization work?
A black-and-white photo still contains the brightness of every point; only the colour is missing. A colourisation model is a neural network trained on many colour photos turned grey, so it learns which colours usually go with which shapes and textures: sky is blue, foliage is green, skin has a certain range of tones. Given your photo it predicts the two colour channels, and the tool combines them with your photo's own brightness at full resolution. The result is a plausible colouring, not a recovery of the original colours.
Are the colours historically accurate?
No, and no automatic colorizer can promise that. Many different colours turn into the same grey, so a model can only choose likely colours. Skies, plants and skin usually come out believable; clothing, cars, painted walls and flags are often wrong or muted. If the true colours matter, for a museum, a book or a family record, treat the result as an illustration and label it as AI-colourised.
Is my photo uploaded to a server?
No. The model runs in your browser with WebGPU or WebAssembly, and the photo stays on your device. The only download is the 16 MB model and the 24 MB AI engine from this site on first use, which your browser then caches.
Does it repair scratches, tears or fading?
No. This tool only adds colour. Scratches, dust spots and creases stay as they are, and a large dark scratch may even pick up a colour of its own. Retouch damage in a photo editor first (a healing brush is the usual tool), then colorize. If the photo is small or soft, upscaling it first with an AI upscaler helps more than colourising at low resolution.
Why do the colours look muted or brownish?
A small model plays safe when it is unsure, and safe means low saturation and warm tones. Raise Colour strength to 120–150% for livelier colour; the tool keeps every colour inside the printable sRGB range, so it will not clip into neon. Scenes the model rarely saw in training, such as illustrations, drawings and X-rays, colourise poorly.
Can I colorize a photo that is already in colour?
You can, but it throws the existing colour away: the tool keeps only the brightness and replaces the colour with its own guess. That can be a fun way to restore a badly faded colour print, but for a normal colour photo the original will be better. The page warns you when a photo already looks colourful.
What happens to your photo, step by step
Knowing the pipeline helps you get better results, so here it is in full:
• The photo's brightness (the L* channel of the CIE Lab colour space) is computed at its full resolution. This is the part of the picture that stays exactly as you uploaded it.
• A 256 × 256 copy of that brightness goes to the AI model, which predicts the two colour channels (a* and b*) at that size. This takes well under a second on a laptop.
• The small colour map is enlarged with a guided filter: in every small area it learns how colour relates to brightness, then applies that relation to the full-size photo. Colour edges therefore follow the real edges in your photo instead of bleeding as a plain blur would.
• Colour strength scales the colour channels, and any colour that would fall outside the sRGB range is pulled towards grey just enough to fit, so hues do not shift.
Because detail comes from your original, a 4000 × 3000 scan stays 4000 × 3000 with all its grain and sharpness. Because colour is predicted at 256 px, very small objects (a tie, an earring, a sign in the distance) usually get the colour of their surroundings.
Tips for better colorized photos
• Scan or photograph the print evenly lit, without glare, and crop away the border and mount; the model reads the whole frame, and a white border shifts its guesses.
• Sepia and toned prints work as they are: only brightness is used, so the brown tint is replaced, not added to.
• Remove large scratches and stains first. The model sees them as objects and may colour them.
• If the result is close but bland, try 120–130% colour strength before anything else.
• For a photo you will print large, upscale a small scan first, colorize second: the colour step keeps whatever detail the input has.
• Keep your original scan. The colorized file is an interpretation and should be stored next to the original, not instead of it.
About the model
The colourisation model is Mini Photo Colorizer v3 by User-2468 on Hugging Face, a 4-million-parameter network (a MobileNetV3 encoder with a palette decoder) distilled from DDColor (Kang et al., ICCV 2023). It is released under the Apache 2.0 licence, which allows this use; the unmodified ONNX file is served from this site with its licence next to it (/models/mini-photo-colorizer-v3.LICENSE.txt).
It was chosen over larger models because it is 16 MB instead of 130 MB or more, so it loads quickly even on a phone connection and fits on this site's own servers. Its authors state its known limits plainly, and we repeat them: muted or warm colour choices, wrong hues for clothing and objects, some colour bleeding around small objects, and weak results on unusual scenes and illustrations.
