Hands holding a faded black-and-white family portrait print beside a laptop showing a softly coloured version of a similar old photograph
AI-generated image, not a photograph of a real person or place.
aifunphotos

WHAT THE RESEARCH SAYS

The colour was
never in the photo.

PLAUSIBLE is not the same as true

  • Colour is a guess
  • Faces get redrawn
  • Keep the original
  • Enjoy the copy, protect the source

AI can colourise a great-grandparent's portrait in seconds.

It is fun, and often moving. But an old monochrome print holds no record of colour, so the software paints in what is plausible, learned from modern photos. Here is what is guessed, and how to keep the real thing.

Updated

Check a photo

THE SHORT ANSWER

Fun to make,
but it is a guess.

AI edits add things that were not recorded.
Three things worth knowing before you share one.

COLOUR

Picked, not recovered

A historian writing in History Ireland puts it plainly: monochrome photos do not contain colour information that can be restored. Colourising software learns its tones from modern photographs.

Check the process

FACES

Detail drawn from other faces

Face restoration research such as GFP-GAN uses what faces generally look like, learned by a face model, to fill in detail a blurry original never captured.

See the research

ORIGINALS

Scan first, edit a copy

The Library of Congress advises keeping the highest-quality version of an important photo and at least two copies, stored in different places.

How to do it

WHAT DID THE CAMERA SEE?

Some old photos
were partly colour-blind.

Early photographic materials did not respond to every colour. Pick a type of emulsion to see how much colour detail was lost before any software ever saw the picture. The bar shows how much is missing.

Emulsion check

Choose a process
Wet plate collodion (mid-1800s)Blue only

Sensitive only to blue (and ultraviolet) light. Greens, reds and yellows can all print as the same tone, so there is no way to tell them apart afterwards.

Orthochromatic (from 1873)Blue and green

Dyes added to the emulsion extended it to green light, but not red. Blue things come out lighter and red things darker than the eye would expect.

Panchromatic (most modern B&W)All colours

Most modern black-and-white films respond to red as well, so tones follow the scene more closely. Even so, a grey tone still says nothing about which colour made it.

Sources: History Ireland, 'Deoldifying Ireland' (on collodion and colour information); Wikipedia, Orthochromasia (on orthochromatic and panchromatic emulsions).

Older processes record some colours as the same grey, so a colouriser has even less to go on than it seems.

TWO NUMBERS WORTH KNOWING

Convincing is
not the same as correct.

One figure on how believable AI colour can be,
one on what it learned from.

32%

Viewers fooled by the AI colour version

In a 'colorization Turing test', people chose between an AI-coloured and a real colour photo. The 2016 method of Zhang, Isola and Efros fooled them on 32% of trials.

COLOURISATION TEST · 2016

1m+

Colour images the system learned from

The same system was trained on over a million colour images. Its job, in the authors' words, is to produce a plausible colour version: a likely answer, not a recorded one.

TRAINING DATA · 2016

The researchers describe colourisation as 'clearly underconstrained': many different colourings fit the same grey photo, so the software has to choose one.

FOUR HABITS TO TRY

Have the fun,
keep the history.

A flatbed scanner with its lid open and an old black-and-white photograph placed face down on the glass, cotton gloves beside it
AI-generated image, not a photograph of a real person or place.

None of this stops you enjoying an AI-coloured family picture. It just makes sure the real photograph survives, and that anyone who sees the new version knows what it is.

  1. Handle the print gently

    The Library of Congress asks its own staff to wear gloves when handling photographs and to avoid feeder scanners for them.

  2. Save the plain scan

    Keep the untouched, highest-quality scan as your master file and run AI tools on a copy.

  3. Name and tag it

    Give files descriptive names and tag the people in them, as the Library of Congress suggests, so the story stays with the picture.

  4. Say it is colourised

    When you share an edited version, say that the colour and any sharpened faces were added by software.

WHY IT MATTERS

Edited copies
travel without notes.

A close-up of an old sepia studio portrait print with fine scratches and silvering at the edges, lying on a wooden table
AI-generated image, not a photograph of a real person or place.

A historian of visual culture notes that colourised images quickly lose any explanation of what artistic licence was taken. Shared online, a guessed colour can start to look like fact.

The same writer points out that colourised versions often smooth skin and add a 'vintage' palette. The original print, with its own process and flaws, is the actual historical record.

Questions

Questions people ask

Can AI recover the real colours of an old photo?

No. A monochrome photo records light and dark, not colour. Colourising software chooses colours that are plausible, based on what it learned from modern colour photographs.

Why do some old photos look odd even before colouring?

Early processes did not see every colour. Collodion plates responded only to blue light, and orthochromatic materials missed red, so reds could print darker than expected.

Is a restored face still the same person?

It is a reconstruction. Face restoration methods such as GFP-GAN fill in detail using a model of what faces generally look like, so fine features may be invented rather than recovered.

How convincing is AI colouring?

In a 2016 study, people shown an AI-coloured and a real colour photo picked the AI version as real on 32% of trials. Methods have moved on since, but they still guess.

What should I keep after scanning?

Keep the highest-quality untouched version, make at least two copies in different places and check them at least once a year, as the Library of Congress suggests.

Sources

Where the facts come from

  1. Zhang, Isola and Efros: Colorful Image Colorization (arXiv, 2016)
  2. Wang et al.: Towards Real-World Blind Face Restoration with Generative Facial Prior (GFP-GAN, arXiv)
  3. History Ireland: 'Deoldifying' Ireland, does photo colourisation bring us closer to the past?
  4. Wikipedia: Orthochromasia (orthochromatic and panchromatic emulsions)
  5. Library of Congress: Personal archiving, digital photos
  6. Library of Congress: Preservation guidelines for digitizing library materials

ABOUT THIS NAME

Interested in leasing aifunphotos.com?

aifunphotos.com is part of the DoodleSale portfolio. It isn't in an auction right now, but you can ask about leasing it.

Ask about this name