AI Colorize for Black-and-White Photos
How does AI turn a black-and-white photo into color?
Add estimated colors to a black-and-white photo with an AI colorization model. Choose the model size and export format, then compare the result with the original. The colors are inferred rather than recovered from historical evidence, so clothing, skin tones or objects may need review. Keep the original scan when preserving a family or archival photograph.
Cost: Free to use. File size and usage limits apply.
How to colorize a black-and-white photo
Upload one photo
Click Upload a black and white image to colorize or drop a JPG, PNG, WebP, BMP, TIFF, AVIF, HEIC or HEIF file of up to 200 MB. Images you uploaded or created in other image tools are listed under Files, so you don't need to upload them again.
Choose High quality or Fast
High quality is selected by default and predicts colors at an internal resolution of 512 × 512 px. Fast uses a smaller model at 256 × 256 px. Both keep the source pixel dimensions; compare their predicted colors for your photo.
Pick the output format
WebP is preselected under Output. WebP and PNG store the result losslessly. Choose JPG when you accept lossy compression to reduce file size; the actual size depends on the image.
Colorize and compare
Click Colorize Image. When the job is done, drag the slider across the preview to compare the Original with the colorized version.
Download or keep editing
Click Download to save the file. The result also appears under Files, so you can open it in another image tool right away, for example to enlarge it with AI Upscale.
What the model changes and what it keeps
The colorizer runs DDColor, an open-source colorization model, on our servers. It splits your photo into lightness and color, keeping the original lightness at full resolution and combining it with predicted colors. It does not apply targeted sharpening or repair dust, scratches or film grain. The output format and conversion to 8 bits per channel can affect fine detail in the saved file.
To predict color, the model looks at a copy of the photo squeezed into a 512 × 512 px square with High quality, or 256 × 256 px with Fast. The predicted colors are then stretched back to the full size of your photo. That is why the output always has the same pixel dimensions as the input, and why color edges are softer than the detail they sit on.
Settings explained
- High quality
- The default. Uses the full DDColor model with a 512 × 512 px color pass. Choose it for final results and for photos with many small objects.
- Fast
- Uses the tiny DDColor model with a 256 × 256 px color pass. The output keeps the original pixel dimensions, with less resolution for placing color around small details and no dedicated sharpening step. Useful for a first comparison before you run High quality.
- Output: PNG, JPG or WebP
- WebP (preselected) and PNG are saved without loss. JPG is saved with lossy compression at quality 95. It can reduce file size, but is not guaranteed to be smaller than both lossless formats for every image. To get another format such as TIFF or AVIF, convert the result with Convert Image.
Getting better colors
Crop to the subject first
The color pass has a fixed size. On a 3000 × 2000 px group photo, High quality predicts one color value for every block of about 6 × 4 px (12 × 8 px with Fast). Crop a 600 × 800 px portrait out of it with Crop Image, and the same pass covers about 1.2 × 1.6 px per value, so colors follow faces and edges more closely.
Check faces and large plain areas
Wrong guesses are easiest to spot in skin tones, clothing, walls and sky. If a result looks off, try the other model: the two can choose different colors for the same object.
Fine-tune the result
If colors come out too strong, too pale or with a cast, correct Saturation, Temperature or Tint in Color Adjust afterwards.
Formats and limits
- Input formats
- JPG, PNG, WebP, BMP, TIFF, AVIF, HEIC, HEIF
- Output formats
- PNG, JPG (quality 95) or WebP (lossless)
- Output size
- Same pixel dimensions as your image, 8 bits per channel. Transparency is not kept.
- Maximum file size
- 200 MB per image
- Files per job
- 1
- Processing
- On our servers with the DDColor model. Your photo is not sent to an external AI service. A job can run for up to 15 minutes.
- Metadata
- Camera data, GPS location and embedded color profiles are not copied into the result.
- Watermark
- None
- Storage
- Uploads and results are deleted automatically after 48 hours for guests and after 30 days for free accounts. Paid plans can keep files longer.
Frequently asked questions
Can I convert a black and white photo to color automatically?
Yes. There is nothing to paint or mark: upload the photo, click Colorize Image, and the model colors the whole image in one pass. You can't pick colors for individual objects. To shift the overall tone afterwards, use Color Adjust.
Are the colors historically accurate?
Not necessarily. The model has learned which colors usually go with which shapes and textures, such as skin, foliage and sky. It cannot know the real color of a uniform, a car or a dress. Treat the result as an interpretation and label it as colorized when you publish or archive it.
Does it work on sepia or faded color photos?
Yes. The model reads only the lightness of each pixel, so a brown sepia tone or faded dyes are ignored and replaced with new colors. The same happens with a color photo: its existing colors are replaced, not enhanced.
Why is the colorized file larger than my scan?
The result has three color channels instead of one gray channel, and PNG and WebP are saved without loss here. A grayscale JPG scan can therefore turn into a much larger file. Try JPG under Output if you can accept lossy compression, then compare its file size and detail with the lossless output.
Will colorizing make my photo sharper or larger?
No. The output keeps the input's pixel dimensions, without targeted sharpening or repair of scratches, dust and blur. Fine details can be affected by the chosen output format. To enlarge the colorized photo, open it from Files in AI Upscale.