That moment I realized AI image generators still can't do hands right

I was messing around with DALL-E 3 last night trying to generate a simple picture of someone holding a coffee cup. Every single attempt had fingers that looked like twisted noodles or extra thumbs growing out of random places. It got me thinking how even with all these massive advances, some basic human anatomy still trips up the models. Has anyone else noticed AI struggling with hands specifically, or is there a technical reason this is so hard for them to figure out?

2 comments

2 Comments

ninas671mo ago

The hand thing is absolutely unhinged. I spent a good hour last week trying to get an AI to generate a picture of me shaking hands with someone, and it just kept giving me these nightmare tentacle things. @julia_anderson nailed it about the training data, I bet there's like 10,000 pictures of a sunset for every one picture of a hand holding a pencil. The models have probably been trained on so many freakish AI-generated hands by now that they think that's just how hands look.

julia_anderson1mo ago

Yeah I spent like TWO HOURS last week trying to get Midjourney to generate a hand pointing at something. It came out looking like a melted crab claw with extra thumbs. It is WILD to me that these models can paint photorealistic landscapes but hands just turn into nightmare fuel. I read somewhere that hands are hard because they have so many tiny moving parts and the AI sees them as separate little objects instead of one connected thing. Plus the training data probably has way more pictures of faces and cars than closeups of people's hands doing specific things.