Colorize black and white photos

How to colorize a photo

1. Add your photos. Drop one or many anywhere on the page, paste one, or click the box to choose files. Scans, phone photos and iPhone HEIC pictures all work.
2. Press Colorize. The first time, the page downloads the colorizing model (about 130 MB) and your browser keeps it, so this happens once. After that each picture takes a few seconds to colorize.
3. Look at the result. The comparison slider shows your original on the left and the colorized picture on the right. Click another picture in the list to compare that one.
4. Adjust the amount of color if you want more or less. The pictures change at once.
5. Download. Use the link on a picture, or Download all as ZIP for a whole batch.

You can add more photos at any time. Your original files are never changed, and each result keeps the original name with -colorized added, for example grandma-1952.jpg becomes grandma-1952-colorized.jpg.

What to expect: the colors are a guess

A black and white photo does not contain its colors. They were lost when the photo was made, so no program can know what they were. The model has learned from a large set of color photographs what things usually look like: the sky is blue, grass is green, skin has certain tones, a brick wall is reddish. It uses that to paint the most likely colors. The result is often convincing, and it is still an educated guess, not the truth.

We tried it on color photos that we turned black and white first, so that we knew the real colors:

• A mountain under a blue sky, with a town below. The sky came back blue, the mountain slopes brownish green and the roofs of the town orange and red. Measured against the real photo, the average error in the color channels fell from 22.1 for plain grey to 13.7, which is much closer. One patch of the sky in a corner came out reddish, which is a typical flaw of colorizing: now and then a region gets a color that does not belong.
• A white cat with orange patches on a dark blue blanket. The blanket came back blue and the orange patches on the cat were found. The white fur came out slightly cream. This photo was almost colorless to begin with, so the colors that were added made it a little less true: the error against the real photo was 5.1, compared with 2.5 for plain grey. When a photo had little color, a gentle amount of color is the more honest choice, which is what the slider is for.

So treat the result as a likely version of the scene. For a family photo, a historic picture or a piece of art, that is often what you want. Do not use it as proof of what color something really was.

The amount of color

The Amount of color slider runs from 0% to 150% and starts at 100%, which is the color exactly as the model made it. At 0% you get the grey picture back. At 50% the color is half as strong, which suits a photo where the model went too far. Above 100% the color is stronger, which can look vivid but also makes mistakes more visible. In our test the colorfulness of the mountain picture was 9.3 at 50%, 18.7 at 100% and 28.0 at 150% on the Lab color scale, rising steadily.

The slider works at once and does not run the model again, because the page keeps the colors the model made for each picture. That also holds for the file type and quality.

Saving the pictures

• JPG is the default, with a quality from 50 to 100 (starting at 92). It makes small files that open everywhere.
• PNG keeps every pixel exactly, with larger files. In our test a 1024 × 768 photo was 157 KB as a JPG and 1.5 MB as a PNG.
• If a picture is larger than about 24 megapixels, it is made smaller before colorizing, because the page holds a few copies of it in memory.

Photos that work best, and tips

• Sharp, well-lit scans work best. Faded, scratched or very dark photos can give weaker colors. The color fixer can improve the contrast and the brightness first, even on a black and white picture.
• Crop away borders and stains first. A black frame or a scanned edge can confuse the colors. Use the crop tool.
• Photos with clear everyday things (people, sky, plants, buildings) tend to work better than unusual objects, because the model learned from many examples of them.
• Check faces and clothes. Skin tones are usually plausible, but a colorful dress or uniform can come out in the wrong color, and that is not something the picture can tell the model.
• Want it sharper or larger? Colorize first, then use the image upscaler.
• Dust, spots or a stranger in the background? Take them out first with the object remover, then colorize.

Photos that already have color

You can add a color photo too. The page turns it grey first and then colors it again, which gives a different version of the colors, not a repair of the original ones. The list tells you when a picture already had some color. A tinted photo, such as a brown sepia print, may count as having some color in this check, and it is turned grey first in any case.

How it works

The model is DDColor (the compact "tiny" version, Apache-2.0 license), from a 2023 research paper on photo-realistic colorization. It looks at the lightness of the picture, scaled to 512 × 512 pixels, and predicts the two color channels. The page then puts that color back onto your original, full-size lightness, so the picture keeps all its detail and only the color is worked out at the smaller size. Because the color is predicted at 512 pixels, it can spill slightly over very thin edges and tiny details.

We checked the page's result against the model's own reference pipeline run natively on a computer, using the same model file. On our two test photos, the colorized pictures differed from it by an average of 0.7 and 1.3 levels out of 255, and the lightness of the result matched the original to within 0.1 on the 0 to 100 lightness scale.

Limits

• Speed. In our test a picture took a few seconds after the model was loaded (about 6 seconds on the test computer). A slower computer or a phone takes longer, and the first picture also waits for the 130 MB download.
• Memory. The model is large (about 130 MB) and is held in memory while you work. If a phone struggles, add one picture at a time.
• Mistakes happen. A region can get a wrong color, and a big patch of one tone can get an uneven color. Try a lower amount of color, or crop to the part you care about.
• It does not repair a photo. Scratches, dust, blur and fading stay as they are. It only adds color.

Privacy

• The photos are colorized inside this tab, and your files on disk are never changed or deleted.
• The model file is downloaded the first time from Hugging Face, a model hosting site, and your browser keeps a copy.
• For what this site collects and keeps, see our privacy policy.