Upscaling and old photos
Lesson 7 of 13 in the course “AI Images and Slides”. Free, with animations and a quiz.
Where the pixels come from
Liya found an old photo of her first batch of candles, but it’s small and blurry. The search “improve photo quality with AI” fixes that in a minute. But it matters to understand what happens. Upscale a photo 4× in width and height and there are 16 times more pixels. So 15 out of every 16 pixels are invented by the AI. It draws what usually appears in similar photos. It mostly looks great, but it’s no longer quite the same shot.
The PULSE case
In 2020 researchers showed PULSE, a model that turned a blurry face into a sharp one. People fed it a pixelated photo of Barack Obama, and it produced the face of a white man. It didn’t miscalculate. You simply can’t identify a person from 16 pixels, so the AI filled in the most typical face from what it learned on. The takeaway is simple: AI doesn’t “reveal” hidden details, it guesses them. So an enhanced photo can’t be used to identify a person or read a licence plate.
Old photos and colour
Restoring grandma’s photo is one of the warmest use cases. AI removes scratches, creases and stains, and can then colorize the photo. But colour has a catch. A black-and-white photo holds no colour information, only lighter and darker. The model knows which colours dresses and wallpaper usually had back then, but not what yours were. So colorizing is a pretty guess, not the truth. Ask relatives if they remember and put the colours into the prompt.
What to use
For old photos GigaChat is handy: it runs Kandinsky 6, which is free and restores pictures on request in Russian. If you just need a bigger image, there’s Upscayl. It’s a free open-source app for Windows, macOS and Linux that works offline, so your photos never leave your computer. ChatGPT and Gemini enhance and edit with words, but may quietly change facial features. And don’t forget your phone’s editor: it often fixes noise and sharpness by itself.
Where not to enhance
Since AI fills in details, an enhanced photo is no good where the truth matters. That’s documents, evidence for insurance or court, news and finding people by their face. If you attach an “enhanced” photo of a crack in a dispute with a repair company, it’s easy to challenge. But a family archive, an avatar, a post or printing a product photo are great uses. In the archive, honestly label it “restored with AI”, so that in 20 years nobody mistakes a guess for a fact.
Restoring step by step
A good result starts with a good source. Scan the photo at 600 dpi or shoot it straight on with your phone, in daylight and without glare. Keep the original separately. First ask to remove the damage, then colorize if you want. Always compare the face to the original: the eyes, the smile and the nose should stay familiar. Copy the prompt below.
Check yourself
- A photo was upscaled 4× in width and height. What share of the pixels did the AI invent?
Correct answer: 15 out of 16. 4 × 4 = 16, so each real pixel gets 15 new ones. - What does the PULSE and Obama photo story teach?
Correct answer: AI guesses details rather than finding the real ones. You can’t identify a person from 16 pixels, so the model drew the most typical face from its training. - Liya wants to attach an “enhanced” photo of a crack in a dispute with a repairman. Should she?
Correct answer: No, the original is better: an enhanced photo is easy to challenge. Evidence needs the truth, and AI may have drawn in details of the crack.