Colorize a Black and White Photo Without Inventing History
How AI colorization works, why the colors are guesses, how to research period details, and how to label a colorized family photo honestly in your archive.
Published
Jump to the editor ↓Short answer: AI colorization is a plausible interpretation, not recovered fact. The model infers color from tone and context, so skin, fabric and walls can come out wrong. Restore damage first, supply any colors you actually know, compare faces with the original, and label the color file as an interpretation while keeping the black-and-white scan as the archive copy.
Colorization adds an interpretation, not evidence. Keep the original scan and say so in the caption.
How the model chooses a color
A black-and-white photograph stores brightness, not color. Every pixel records how much light reached the film, and the color information was never captured. To colorize, the model reads shapes, textures and tonal values, then assigns colors that fit what it has learned from other photographs. Skin is usually warm, foliage is usually green, and a dark uniform is often navy or black.
That is inference, not recovery. The model does not know what dress your grandmother wore, which regimental facing a uniform had or whether the kitchen wall was cream or pale blue.
What it usually gets right, and where it guesses
| Area | Typical result | Risk |
|---|---|---|
| Skin | Plausible in a clean, well-lit scan | Tones shift in dark or overexposed areas, and across different skin tones |
| Sky and foliage | Often convincing | The season, time of day and weather are unknown |
| Clothing | A believable generic color | A specific dress, team color or uniform can be wrong |
| Interior walls | A soft neutral or period color | The real paint color is unknowable from the print |
| Metal and insignia | Sometimes plausible | Fine detail is easy to misread |
| Small objects | Weakest area | Low resolution leaves too little information |
The pattern is simple: the more specific the object, the less the model can know. A family's tartan is not a field of grass.
Research the period before you colorize
A few minutes of research turns a guess into a documented choice. Start with:
- relatives who remember the person or the place
- other photographs of the same event, room or uniform
- reference images from local historical societies, museum costume collections or military archives
Write down what you learn and where you learned it. "Navy dress, according to my aunt who owned it" is a real source. "It looked right" is not. When you have no source, keep the palette muted and say in the caption that the colors are estimated. The colorize photo tool is designed for this kind of careful, documented interpretation.
Restore first, colorize second
Cracks, dust, fading and torn edges confuse a color model. A scratch across a face can be read as a shadow, and a faded patch can attract a color that was never there. Repair the damage first with the old photo restorer, then colorize the cleaned file.
This order also protects the archive. Keep the restored black-and-white version as the master copy and treat the colorized file as a separate derivative, so a questioned color never replaces the honest version.
Prompting with known colors
If you know a color, say it plainly. Specific instructions beat vague ones such as "make it colorful". A useful prompt names the palette, protects the person and rules out additions:
Colorize this black-and-white photograph with muted, period-appropriate colors. The dress is navy blue and the wall is pale green. Preserve the original faces, clothing shapes, grain and composition. Do not change the person's identity, age or expression. Do not add or remove objects. No text or watermark.
When you do not know the colors, ask for a restrained, natural palette rather than vivid saturation. Strong colors make the guess look more certain than it is. You can also colorize in stages, correcting one area at a time instead of accepting a fully saturated result.
Check skin tones and faces
Skin is the first thing a viewer notices and the hardest area to get right. Compare the colorized face with the original at the same size. Look at the forehead, cheeks, hands and the shadows under the chin; a tone that is too orange, too pink or too gray reads as artificial even when the rest of the image looks fine.
Colorization can also drift the likeness. The model may smooth skin, change the shape of an eye or invent detail in a dark area. Check the eyes, nose, jaw, hairline and any distinguishing mark. If the person no longer looks like themselves, regenerate from the original scan with a stricter instruction rather than layering another edit on top.
Label the file and protect the archive
A colorized photograph is an interpretation, and the caption should say so. A short line such as "Colorized interpretation, colors estimated" is enough for a family album, a slideshow or a framed print. For a book, an exhibition or a historical display, be more specific about what was known and what was guessed.
Keep the original scan untouched under its own filename, and save the colorized version separately, for example with a -colorized suffix. Do not overwrite the only copy. If you share the image online, include the label in the post, not only in a reply thread where most people will never see it.
Trying it on a real photo
EzCreate accepts JPG, PNG, WebP and AVIF uploads up to 10MB. You can open the editor and upload a photo for free; sign in to claim the daily preview credits. Larger runs, downloads and 2K or 4K output use purchased credits. Results are deleted automatically after 7 days, and EzCreate does not add its own watermark.
The model picker includes GPT Image 2, Nano Banana 2 and Nano Banana Pro, Seedream 4, 4.5 and 5 Lite, Flux 2, Qwen, Ideogram and Grok Imagine. Colorization is a judgment-heavy task, so run the same scan on two models and compare faces, fabric and background before you choose. If you need to extend the frame around a portrait, the AI image extender can add room without stretching the face, but the new area is generated too and should be labeled.
Frequently asked questions
How accurate is AI colorization?
It is plausible, not factual. The model infers color from tone and context, so a dress, wall or uniform can be wrong. Supply known colors and label the result as an interpretation.
Should I colorize before or after restoring?
Restore first. Repairing cracks and fading gives the color model a cleaner image, and you can keep the restored black-and-white file as the archive copy.
Can I correct a wrong color?
Yes. Describe the correct color in the prompt and regenerate, or correct the result in a photo editor. Start from the original scan each time.
Will colorization change the person’s face?
It can, especially in low-resolution areas. Compare the result with the original and regenerate if the likeness drifts.
Can I colorize a photo that is still damaged?
You can, but the damage will influence the colors. Repair the obvious cracks and fading first, then colorize the cleaned file so a scratch is not mistaken for a shadow or a stain for a garment.
Does colorization work on a very small scan?
It runs, but the model has less real detail to work with and will invent more. Rescan the print at a higher resolution if you can; a larger source is the most reliable way to improve both the color and the likeness.
Try EzCreate now
Start with a scan you can afford to experiment on, restore the obvious damage, and colorize with a muted palette. Compare the face with the original before you share it, and keep the black-and-white file as the record.