AI Photo Colorizer — Pro, on our server
The model predicts chroma and never touches luminance. That single sentence explains most of what you will see. A black-and-white photograph is a record of brightness with the colour information discarded, and many different colours collapse to the same grey on the way. DDColor takes the brightness data as fixed, segments the frame, forms an opinion about what each region depicts, and outputs the colour values most likely to have produced those greys in a real photograph. Your original structure — focus, grain, contrast, composition — survives untouched, which is why a soft photograph comes back soft and a damaged one comes back damaged, in colour. It is also why the result is a plausible photograph rather than a verified one: the model is reconstructing a distribution, not recovering a fact.
Know in advance which regions to trust. Four categories are consistently reliable, because the physical world constrains them: vegetation, skin under reasonable light, common building materials, and sky. Four are consistently unreliable, because a person chose them: clothing, paint, vehicles and anything printed or heraldic. Pale eyes fall in the second group — blue, green and grey irises all photograph to a similar mid-tone, and the model reaches for the commoner brown, which is precisely what the demo above shows on a blue-eyed cat. Read a result region by region rather than as a single verdict. Accepting the hedge and repainting the dress is not a workaround; it is how colorization has always been done, with the model taking over the ninety per cent that is drudgery.
It does exactly one job, and the neighbouring jobs have their own tools. It will not sharpen, denoise, resize, repair damage or invert a negative — hand it a negative and you get a faithfully colorized negative. Output is a lossless PNG at exactly the resolution you supplied, up to the 24-megapixel and 25 MB limits. When a photograph needs more, the other Cloud HD models are separate and explicit: the face restorer uses RestoreFormer++ to bring back facial detail in old, blurry or faded photographs and can chain a colorize step onto the end; the photo enhancer cleans blur and noise at the same output size; the object remover paints a brush-masked region back in; and the upscaler is free on your own device with a Cloud HD ×4 mode for Pro. For scans specifically, the old photos page covers digitisation, and the vintage prints page covers film-era quirks.
Where it runs, what it costs, and what is kept. DDColor is far too heavy for a browser, so colorizing is one of seven Cloud HD tools that execute on our GPU rather than on your device — the deliberate exception to MiniPx running everything locally. It is Pro-only and metered: 150 runs a month on Pro Monthly ($3.24 the first month, then $4.99), 200 on Pro Annual ($29 the first year, then $39), prices excluding tax, resetting on the 1st with no rollover, capped at 30 runs an hour, and a failed run is never charged. Your photograph is uploaded over an encrypted connection for that run, processed in the United States, deleted when the run finishes and in all cases within 24 hours, and never used to train anything. Cancelling or changing a plan is handled by a person — email contact@minipx.com — and the plan detail is on the pricing page.