Old film did react in certain ways to specific colors, it wasn’t a 100% accurate grayscale version of the actual image due to the chemical properties of the film. That said, it’s really just a lot of educated guessing. Leaves are green, water is blue, human skin color falls into a specific range of shades, but is that shirt red or blue? Unless there are more specific clues, it’s basically just a guess.
They don’t. They make a guess. Sometimes better, sometimes worse (though of course that thread is four years old and AI has improved since, and also the OOP may have been picking the worst examples to make a point)
Colorizing apps don’t know the real colors they just use AI trained on thousands of colorful photos to guess what colors should go where based on what similar things usually look like. It’s a fancy, digital guessing game with crayons.
The long answer: They don’t know the true original colors, but the algorithms pick up that a large uniform area of one shade near the top third of the image tends to be blue. Large areas with vertical-running shadows near the bottom are green, etc. This is fine for landscapes or very very basic manmade objects which carry very strict patterns. But once you show anything it doesn’t already “know” it breaks down. In reality it might have been very reflective so should show sky blue, but the colorizer doesn’t know that, so it either guesses at random or doesn’t touch it at all.
Comments
Certain things tend to have certain colors.
Grass and leaves are typically green.
Skin tones have their own specific range of colors.
In addition certain clothes had certain colors (such as uniforms often being grey or blue).
Some colors may simply be guessed at using a neutral color, so it doesn’t stick out, even if it is the “wrong” color.
Old film did react in certain ways to specific colors, it wasn’t a 100% accurate grayscale version of the actual image due to the chemical properties of the film. That said, it’s really just a lot of educated guessing. Leaves are green, water is blue, human skin color falls into a specific range of shades, but is that shirt red or blue? Unless there are more specific clues, it’s basically just a guess.
They don’t. They make a guess. Sometimes better, sometimes worse (though of course that thread is four years old and AI has improved since, and also the OOP may have been picking the worst examples to make a point)
Colorizing apps don’t know the real colors they just use AI trained on thousands of colorful photos to guess what colors should go where based on what similar things usually look like. It’s a fancy, digital guessing game with crayons.
The short answer: They don’t
The long answer: They don’t know the true original colors, but the algorithms pick up that a large uniform area of one shade near the top third of the image tends to be blue. Large areas with vertical-running shadows near the bottom are green, etc. This is fine for landscapes or very very basic manmade objects which carry very strict patterns. But once you show anything it doesn’t already “know” it breaks down. In reality it might have been very reflective so should show sky blue, but the colorizer doesn’t know that, so it either guesses at random or doesn’t touch it at all.