The blind spot people miss when training image generators on faces

I was playing around with Stable Diffusion last week and noticed that everyone keeps feeding it the same three types of face angles. Straight on, three quarter, and profile. That's it. But real photos of people have heads tilted up or down, shots from below, and weird candid angles where someone is turning mid laugh. I tested this by making a dataset of 200 random family photos I took at a birthday party in Denver. The model kept making everyone look like they were standing for a driver's license picture. It matters because if you are trying to generate natural looking people for a game or an ad, they need to look like they are actually moving through the world. Has anyone else run into this problem or figured out a good fix for it?

3 comments

3 Comments

jordan_hill1d ago

Is this really that big of a deal for most people?

beth_park19h ago

Oh wait, you haven't dealt with this yourself yet? Because yeah, it really is that big of a deal. I remember when it happened to me and I thought I was overreacting until I talked to like five other people who went through the same thing. It messes with your head more than you'd expect, little things just start piling up and you don't even realize how much it's affecting you until you're already overwhelmed. I felt so stupid for being upset about it at first, but now I know it's totally normal.

the_jake1d ago

Take it from someone who's been there, yeah it's a pretty big deal.