Jason Antic

Here's a shocker: Yes, I'm aware of the excessive purple and other misplaced colors in DeOldify. There's no shortage of people who can't wait to inform me about this. Either make a better model. Or just carry on and trust me that I've thought plenty about it over two years..

The consistency of it is useful, and sometimes striking; so I'm a fan. Outside of time travel, there's nothing you can do to determine what the 'real' colors were, anyhow.

Tell that to the "ideas people" lol

I'm sure you have thought about it. But I am curious: why can't you simply tell it "reduce the amount of purple across the board"? What am I misunderstanding?

1/ So a way to translate what you're saying there into an implementation would be to add a "cost" to adding "too much purple" as a numerical penalty. But the big problem with this is that penalizing with this magic number will inevitably have bad "side effects" because there's

2/ no way to accurately weigh this with a "one size fits all" approach. You want the machine to learn how much is too much purple instead, according to nuanced contextual information. This is what I view as the chief advantage of GANs (which I don't use currently).

3/ To elaborate on what those bad side effects would be: The model would just pick another default color. Much like a game of whack-a-mole!

I'm not sure what color space you use but if you go to something like CieLAB, does the default color become more grayish?