No. 06 — AI Tools · Cloud HDPro · on our GPU

An old portrait, skin tones and all.

A portrait is the hardest thing to colorize well and the most rewarding when it works, because everyone looking at it already has an opinion about the face. Upload the scan; our Cloud HD server returns it in plausible colour at the same resolution, keeps the file only for that run, and deletes it on completion.

Pro · Cloud HDDeleted on completionUp to 25 MB · 24 MP
Black-and-white photograph of a long-haired tabby cat, before colorization
Sample — processed with ddcolor-modelscope 8565d1f7c686
Colorize a portraitCloud HD · our GPU

How it works

  1. Upload the portrait scan with its background intact — up to 25 MB and 24 megapixels.
  2. DDColor infers skin, hair, clothing and backdrop on our GPU, then the upload is deleted.
  3. Download the same-size PNG and hand-correct complexion or eyes if needed. One Cloud HD run.

This page is the photo colorizer, tuned to faces — skin, hair, eyes and the studio lighting that shaped them.

All Image Tools

AI Tools

Solutions by use case

Colorize Old PortraitsPro, on our server

Skin is where colorization is judged, and where it is least certain. Complexion spans a narrow band of hues and a very wide band of luminance, and black-and-white film records only the second of those. DDColor responds by predicting a warm mid-tone that suits the light it can see, which is convincing on a well-lit frontal portrait and drifts on anything shot with hard side-light, heavy studio powder or a strong backdrop reflection. Nobody in your family will notice a slightly wrong wallpaper; everybody will notice a slightly wrong grandmother. Plan on accepting the model’s whole-frame work and adjusting the face by hand — it is minutes of correction on top of a complete base layer, rather than hours of manual colorization from grey.

Studio conventions left fingerprints that the model has to read around. Cabinet cards and carte-de-visite portraits were made with painted backdrops, tungsten or magnesium light, and negatives that were routinely pencil-retouched to smooth skin and open the eyes. All three flatten local contrast, and flat regions are precisely where a colorizer becomes cautious — it will not invent a bold colour where the tones give it no reason to. That is why heavily retouched faces can come back slightly waxy while the sitter’s jacket and the backdrop look excellent. Give the model the whole card, not a tight crop of the head: shoulders, collar and backdrop are context it genuinely uses, and you can always tighten the crop afterwards with the on-device crop tool.

Order the two models correctly. If the eyes and mouth have gone soft, colour will not rescue them — colorization changes chroma only and leaves every bit of the light-and-dark structure exactly as it found it. Run the face restorer first, which uses RestoreFormer++ to bring facial detail back in old, blurry or faded photographs, and take its optional colorize step so the whole thing happens in one pass; that chained job counts as an HD run, of which each plan includes ten a month. If the portrait is already crisp, do not bother — colorize alone and spend the run on the next photo. For a set of portraits from the same studio sitting, the family album page covers keeping a batch consistent.

A face is personal data, so here is exactly what happens to it. Colorization is a Cloud HD tool and therefore the one moment your image leaves your device: it travels encrypted, is processed in the United States on Cloudflare R2 storage and a Modal GPU, and is deleted when the run finishes and in all cases within 24 hours. It is never used to train anything. The model that runs is DDColor and the tool tells you which model produced your file — the model can change, and you can always see which one ran. Cropping, converting and compressing all stay on your own machine, and the free tier never touches the server at all.

How it works

  1. Scan the portrait at high resolution: Cabinet cards and carte-de-visite prints are small — 600 DPI or better keeps enough face detail for the model to work with.
  2. Restore first if the face is soft: Run the face restorer when eyes and mouth have lost definition; it can colorize in the same pass as an HD run.
  3. Colorize the portrait: Upload the JPEG, PNG or WebP, up to 25 MB and 24 megapixels. DDColor returns a PNG at the same resolution.
  4. Correct skin and eyes by hand: Adjust the two regions the model cannot know — complexion and eye colour — in any editor, using the AI output as the base layer.

Frequently asked questions

Why did the skin tone come out wrong for my ancestor?
Because grey does not encode complexion, and studio lighting flattens what little signal there is. The model predicts a warm tone that fits the luminance it sees, which lands well on some faces and noticeably off on others — particularly under strong side-lighting or heavy powder. It is the region most worth correcting by hand afterwards.
Will eye colour be right?
Rarely, unless the eyes are dark. Pale blue, green and grey eyes all photograph to a similar mid-grey on black-and-white film, and the model picks the commonest reading, which is usually brown or hazel. The demo above shows exactly this: a blue-eyed cat comes back amber-green. Fix eyes manually if they matter to the family.
Do retouched negatives confuse it?
Sometimes. Portrait studios routinely pencilled over negatives to smooth skin and lift eyes, which leaves flat, unnaturally even areas on the print. DDColor reads flat areas as low-information and colours them cautiously, so a heavily retouched face can return slightly waxy. Nothing is wrong with the file — that is the retoucher’s hand showing through.
Should I restore the face before colorizing, or after?
Before. Colour reads better on recovered detail than on blur, and the face restorer has an optional colorize step so both happen in one pass — which counts as an HD run rather than a plain one. If the portrait is already sharp, skip restoration entirely and colorize alone; there is no benefit to running a model the photo does not need.
How close should I crop a portrait before uploading?
To the mount, not to the chin. Give the model the shoulders, background and any visible clothing, because it uses the surrounding context to decide how warm the face should be. Cropping tight to a head removes exactly the evidence it reasons from, and you can always crop the coloured result afterwards.