What about people who can look at something and then draw it?
https://en.wikipedia.org/wiki/Stephen_Wiltshire
Do they have to recall specific areas, or do they perceive the entire thing as a fully instantiated mental image.
Glad that you used this exact example! This guy doesn’t have a photorealistic memory. At least it’s far from as good as it’s claimed to be. He’s an artist proficient in a particular style - better than most, but not superhuman.
When he’s not drawing from a direct reference, he’s simply making up details based on assumptions, not on photorealistic memory. Here’s a good example:
https://www.youtube.com/watch?v=FyPqQIHkasI
He looks at a city and then draws a picture of it. It’s very detailed, so we assume he remembered all of it and recreated it accurately.
But if you compare any part of it it to the actual photo of the city he saw, you’ll see that he only recreated it roughly — some landmarks, the general shape of the coastline. He probably got the number of bridges right.
But you couldn’t use this as a map. If you were trying to find a particular building that isn’t among the top 15 most memorable ones, it’s probably not in his drawing, with a completely random building taking its place instead. Every part of that drawing is filled with mistakes and assumptions that would never be made by someone who could actually see the landscape in their mind like a photo.
And it’s the same with every other claim of photorealistic memory - it’s always some kind of trick where people have a decent but realistic level of memory. And then they fill the gaps with tons of generated detail that we either can’t check, or wouldn't bother to check.
This is called building your 'catalogue' in art, especially concept art. In order to draw something (well) from imagination, you should draw it from reference many times. Then when you draw from imagination, your brain will pull from what it knows. And since you studied the subjects, the textures, the shapes, etc, so well, you will have that stored away and will be able to do so.
Yeah, it resembles what you'd get when using gpt 4o for image editing. Of the parts that should have been unaltered, the broad lines are correct, but the exact details are made up. A modern white chair is replaced by some other white chair. A book is replaced by some other book. Etc, etc.
Both brains and gpt appear to be doing lossy compression based on preexisting world knowledge.
