# How to colorize black and white photos with AI (and what to expect)

> A practical, step-by-step guide: prepare the scan, choose between colorize-only and restore + colorize, check the result honestly, and fix the colors that are only guesses.

Published October 1, 2026. By the ColorizeOldPhotos editorial team. Source: https://colorizeoldphotos.com/blog/how-to-colorize-black-and-white-photos

**How we know:** The examples in this guide are public-domain photographs from the Library of Congress that we ran through our own tool in September 2026. We show results exactly as the tool returned them, including colors that look wrong.

Colorizing a black and white photograph takes about half a minute with a modern AI model. Getting a result you can trust takes a little more care. This guide walks through the whole process and is honest about what the technology cannot know.

> **The short version.** Scan the print well, pick **Colorize only** for clean photos and **Restore + colorize** for damaged ones, compare the result with the original, then treat every color you cannot verify as a guess.

## 1. Start with the best source you can get

The model can only work with the pixels you give it. A sharp, flat scan gives noticeably better skin tones and fabric detail than a phone photo taken at an angle. Our [guide to scanning old photos](https://colorizeoldphotos.com/blog/how-to-scan-old-photos) covers resolution and handling; the short version is to scan flat, at 300 to 600 ppi, and to keep the original print.

If you only have a photo of the print, shoot it in even daylight with no flash, hold the camera parallel to the print, and crop out the table or album page. Dark borders can push a colorization model toward odd color casts.

## 2. Decide what the photo needs

There are two kinds of work hiding behind the word “colorize”. Adding color is one job. Removing scratches, stains and fading is another. They use different technology, so choosing the right mode matters.

*Choosing a mode on ColorizeOldPhotos. The same logic applies to any tool.*

| Your photo | Choose | Why |
| --- | --- | --- |
| Clean, sharp, just grey | Colorize only | A dedicated colorization model adds color and leaves the picture as it is. Fast and faithful. |
| Scratches, creases, stains or heavy fading | Restore + colorize | A generative editing model removes the damage first, so color isn’t painted onto cracks. |
| Damaged, but you want to keep it black and white | Restore only | Repairs the damage and keeps the original tones. |

Our colorize-only mode uses DDColor, a research model published at ICCV 2023 by Kang and colleagues. It predicts a color for each region from what it has learned about the world, using two decoders: one for the structure of the picture and one for color.

## 3. Run it and read the result like a skeptic

*Figure: Colorized, Street scene on West Main Street, Norwich, Connecticut (1940), before/after slider at https://colorizeoldphotos.com/gallery. Original: Library of Congress, Prints & Photographs Division, FSA/OWI Collection, public domain.*

Use the before and after slider, then check the same things every time:

- **Skin and hair.** These are usually the most reliable part, because faces are the thing models have seen most often.
- **Sky, grass, wood, stone.** Natural materials have a narrow range of plausible colors, so they tend to look right.
- **Clothes, cars, signs, flags.** These can be almost any color. Treat them as a guess unless you can verify them.
- **Edges and faces.** Colorize-only should not change shapes. If something looks redrawn, you are comparing against the wrong file.

In the street scene above, the road, brick and skin tones read as plausible, while the colors of the coats and the hydrant are anything the model happened to choose. Nothing in a black and white photograph says whether that hydrant was red or yellow.

## 4. Fix the facts that matter

If a color is important to the story of the photo, such as a uniform, a dress or a flag, find a source rather than accepting the model’s guess. Ask relatives, check museum or archive records for that unit or period, and look at real color photographs from the era. Our article on [how accurate AI colorization really is](https://colorizeoldphotos.com/blog/is-ai-colorization-accurate) explains where to look.

## 5. Save, label and keep the original

- Download the result and keep your original scan untouched. The National Archives advises keeping originals after digitizing, because digital files have their own preservation risks.
- Name files without spaces or symbols and record who, what, where and when, as the National Archives suggests.
- Follow the 3-2-1 rule for backups: three copies, on two kinds of media, one off-site.
- When you share the colorized image, say that it is colorized. Anyone who sees it should know the colors are an estimate.

## How the main approaches compare

*Our assessment from testing, not a benchmark.*

| Approach | Speed | Control | Risk to the original look |
| --- | --- | --- | --- |
| By hand in an image editor | Hours per photo | Highest: every color is your decision | Low, if you are careful |
| Dedicated AI colorizer (colorize only) | Seconds | Low: you accept or re-run | Low: structure is kept |
| General image model or chatbot | Seconds to a minute | Some, through prompts | Higher: the whole picture is regenerated, so faces can change |

None of these is always best. For a portrait of a real person you want to recognize, we prefer the approach that changes the fewest things.

> **Disclosure.** We make ColorizeOldPhotos, which is one of the tools mentioned here. The steps above work with any tool.

## Sources

1. [U.S. National Archives, “Digitizing Family Papers and Photographs”](https://www.archives.gov/preservation/family-archives/digitizing.html)

2. [Federal Agencies Digitization Guidelines Initiative (FADGI), Technical Guidelines for Digitizing Cultural Heritage Materials](https://www.digitizationguidelines.gov/guidelines/FADGI_Still_Image_Tech_Guidelines_2016.pdf)

3. [Kang et al., “DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders” (ICCV 2023)](https://arxiv.org/abs/2212.11613)