Your AI-generated Art

OIG.zD9gPy9qV2OgtPpWuW5z
 
animal531 said:
It's not that easy to retrain your model without them, instead I think it rather has to do with training data sets.

Take for example prompts of Keanu R vs Scarlett J riding a motorcycle, with no other input:

View attachment 1621577
View attachment 1621581

He is very much Keanu, but the girl could be almost anyone.

My opinion on why AI is still struggling with some actors (especially actresses or men who change their look often) is this:
1. Keanu has always looked mostly the same in his appearances, same few hair styles, colours, clothing etc. Scarlett's is all over the place
2. The same goes with age, we have say training data for both from ages 18-5x. Keanu has changed over that time but its pretty linear, Scarlett today often looks a lot different from when she was 18 (depending on her makeup, current Botox etc.)
3. So when I tell the AI to make me a pic of Keanu on a bike it is using neural network weights for him that are very close overall, so he always looks like himself. When blending Scarlett her weights are much bigger/wider/random so she becomes very generic, like when you take for example photos of 12 different actresses and blend their faces together, ie. you get someone that's pretty, but very generic looking.

Hands used to be more of an issue because the AI doesn't know that people have 5 fingers per hand, so including a lot of training data where you are showing less fingers confuses it more. You can fix this problem with very specific training that changes the weights towards more realistic ones but it takes specific training.

If they wanted to they could do the same with Scarlett, but obviously that's against the Hollywood/AI game at the moment so there's no incentive unlike with hands.
Click to expand...

animal531 said:
It's not that easy to retrain your model without them, instead I think it rather has to do with training data sets.

Take for example prompts of Keanu R vs Scarlett J riding a motorcycle, with no other input:

View attachment 1621577
View attachment 1621581

He is very much Keanu, but the girl could be almost anyone.

My opinion on why AI is still struggling with some actors (especially actresses or men who change their look often) is this:
1. Keanu has always looked mostly the same in his appearances, same few hair styles, colours, clothing etc. Scarlett's is all over the place
2. The same goes with age, we have say training data for both from ages 18-5x. Keanu has changed over that time but its pretty linear, Scarlett today often looks a lot different from when she was 18 (depending on her makeup, current Botox etc.)
3. So when I tell the AI to make me a pic of Keanu on a bike it is using neural network weights for him that are very close overall, so he always looks like himself. When blending Scarlett her weights are much bigger/wider/random so she becomes very generic, like when you take for example photos of 12 different actresses and blend their faces together, ie. you get someone that's pretty, but very generic looking.

Hands used to be more of an issue because the AI doesn't know that people have 5 fingers per hand, so including a lot of training data where you are showing less fingers confuses it more. You can fix this problem with very specific training that changes the weights towards more realistic ones but it takes specific training.

If they wanted to they could do the same with Scarlett, but obviously that's against the Hollywood/AI game at the moment so there's no incentive unlike with hands.
Click to expand...
with midway could you make wicks hair white?
 
Someone do Karlach from Baldur's Gate 3 throwing a gnome at an elf
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X