How to Colorize Black-and-White Photos with AI: A Complete Guide

How AI colorization works

AI colorization models are trained on millions of pairs of color photographs and their grayscale equivalents. The model learns that certain grayscale tonal patterns correspond to specific colors. Bright, high-contrast areas at the top of an outdoor scene are almost certainly blue sky. Medium-gray organic textures at ground level are likely grass or foliage. Specific tonal ranges in faces correspond to skin tones that vary by ethnicity and lighting conditions.

The seminal research by Zhang et al. at UC Berkeley showed that deep neural networks could produce plausible colorizations by treating the task as a classification problem. For each pixel, the model predicts which of 313 color bins is most likely given the surrounding grayscale context. Modern models have built on this foundation with larger training datasets, higher resolution output. Better handling of ambiguous cases like clothing, vehicles, and interior furnishings where many colors are equally plausible.

Understanding this process explains both the strengths and limitations of AI colorization. The model excels at natural elements — sky, water, vegetation, skin — because these have strong correlations with grayscale values. It struggles with arbitrary choices — the color of a dress, a car, a painted wall — because the grayscale value provides no definitive answer. In these ambiguous cases, the model defaults to the most statistically common color in its training data. Is why AI-colorized clothing often trends toward muted blues, browns, and grays.

• Models learn grayscale-to-color mappings from millions of color and grayscale image pairs.
• Natural elements like sky, vegetation, and skin tones colorize most accurately because they have strong tonal correlations.
• Arbitrary colors — clothing, vehicles, painted surfaces — are educated guesses based on statistical frequency in training data.

Scanning and digitizing your originals

The quality of your colorized result depends fully on the quality of your source scan. A 600-DPI flatbed scan captures fine detail that the AI uses to distinguish between materials. The weave of a wool coat versus the smooth surface of a leather jacket, the texture of brick versus plaster, individual blades of grass versus a gravel path. At lower resolutions, these distinctions blur together and the AI has less information to work with.

If you do not own a flatbed scanner, many public libraries and photo shops offer scanning services. As a last resort, you can photograph a print with your smartphone, but the result will be lower quality. Place the print on a flat, dark surface in even, indirect daylight. Near a north-facing window on an overcast day is ideal. Position your phone directly above the print, parallel to the surface, to avoid perspective distortion. Avoid using flash, which creates specular highlights on glossy prints that the AI will interpret as white objects.

For negatives and slides, a dedicated film scanner produces vastly better results than photographing them against a light source. If you have a box of negatives, a batch scanning service can digitize them at high resolution for a few cents per frame. A worthwhile investment given that each frame is a unique, irreplaceable image.

Repairing damage before colorization

Old photographs accumulate damage over decades. Scratches from handling, foxing spots from moisture exposure, stains from improper storage, creases from being folded or stored without acid-free materials, and overall fading from light exposure. These artifacts must be repaired before colorization because the AI will treat them as part of the image content. A brown foxing spot might be colorized as a facial blemish. A crease across a sky might prevent the model from recognizing that area as sky at all.

Magic Eraser handles most common damage types. Brush over scratches, spots, and stains — the AI reconstructs the underlying grayscale tones from surrounding context. For deep creases that have cracked through the emulsion, make multiple passes, allowing the AI to rebuild the area gradually. For large areas of damage — water stains that cover a quarter of the image, torn edges, completely faded sections — the AI performs best when you work from the edges of the damage inward rather than trying to fix the entire area at once.

Fading is a separate problem from physical damage. A uniformly faded photograph has lost contrast. Blacks have shifted to dark gray, whites to light gray, and the midtone range has compressed. AI Enhance recovers this lost tonal range by stretching the histogram back to full black-to-white contrast. Run this boost after physical damage repair but before colorization. The colorization model needs strong contrast to distinguish between different surfaces and materials.

• Repair physical damage — scratches, spots, stains, creases — with Magic Eraser before colorization.
• Work from the edges of large damaged areas inward for best AI reconstruction results.
• Restore faded contrast with AI Enhance after damage repair but before colorization.
• Unrepaired damage artifacts will be colorized as if they are part of the image, creating permanent errors.

Optimizing images for the best colorization results

Between damage repair and colorization, a preparation step greatly improves results. Run the cleaned image through AI Enhance to optimize contrast, sharpen soft details. Lift areas that are too dark to distinguish. The goal is not to make the black-and-white photo look better for its own sake. It is to give the colorization model the clearest possible input.

Pay attention to areas where the subject blends into the background. In many old photographs, a person's dark hair merges with a dark background, or a white shirt disappears into a bright sky. If the AI cannot see the boundary between subject and background, it cannot colorize them separately. AI Enhance often recovers enough edge detail to separate these areas. If it does not, consider using Magic Eraser to subtly darken or lighten the boundary area before colorization.

Cropping is also worth considering. Old photographs often have damaged borders, uneven edges from printing, and wide margins. Cropping to the actual image content removes these distractions and gives the AI model a cleaner composition to analyze. Save the cropped version as a separate file so you preserve the original scan with its full borders for archival purposes.

Reviewing and refining colorized results

AI colorization is impressive but imperfect. Every colorized image deserves a careful review before sharing or printing. The most common issues fall into predictable categories that are easy to spot once you know what to look for.

Skin tones are the most critical element to check. The AI determines skin color from a combination of grayscale value, facial feature recognition, and contextual clues. In most cases it is accurate, but unusual lighting. Strong backlighting, theatrical stage lighting, or heavy shadow — can shift skin tones toward unnatural hues. If skin looks too pink, too yellow, or too gray, a targeted color temperature adjustment brings it back to a natural range.

Clothing and fabric is where the AI guesses most often. If you know that your grandmother's wedding dress was ivory rather than white, or your grandfather's military uniform was olive drab rather than the navy blue the AI chose, you can use AI Filter adjustments to shift the hue in those specific areas. Historical reference photos from the same era can guide these corrections.

Seasonal context is another area to verify. The AI does not know what month a photo was taken. It may colorize autumn trees as summer green or winter grass as spring emerald. If you know the season, adjust foliage and landscape colors accordingly.

• Check skin tones first — unusual lighting in the original can cause unnatural color shifts.
• Verify clothing colors against family knowledge or period reference photos.
• Correct seasonal context — the AI defaults to summer-like foliage regardless of when the photo was taken.
• Run a final AI Enhance pass after manual corrections to unify overall color temperature.

Preserving and sharing colorized family photos

A colorized family photograph is a meaningful gift and a powerful connection to family history. The Library of Congress recommends saving archival images in lossless formats. PNG or TIFF — at the highest resolution available. Save three versions of each photo: the raw scan, the restored black-and-white version, and the colorized version. This archival approach means you can re-colorize the image in the future as AI models continue to improve without starting from scratch.

For sharing, resize and export as high-quality JPEG. Print services accept JPEG at 300 DPI. A 4000-pixel-wide image prints beautifully at 13 inches across, large enough for a framed display. For digital sharing in family group chats and social media, the full-resolution file works as-is on modern platforms.

Consider creating a before-and-after comparison — the original black-and-white on the left, the colorized version on the right — which makes the change right away visible and often sparks conversations about family history that the photo alone does not. These side-by-side images are mainly popular for social media posts honoring grandparents, commemorating anniversaries, and celebrating family milestones.

• Save raw scans, restored grayscale, and colorized versions separately for future re-processing.
• Export at 300 DPI for printing and full resolution JPEG for digital sharing.
• Before-and-after comparison images spark more family engagement than colorized photos alone.