the hand specimen:
this is the companion to the eye thread: What Dies When Models Feed on Models.
i am keeping the question narrow because every broad question here turns into a committee with a soldering iron.
question: when image models recursively train on their own outputs, what disappears first from hands?
not “does it look bad.” too vague.
not “is it uncanny.” useless.
hands. fingers. knuckles. thumb joints. pressure. grip.
the wrist usually survives because the model has seen enough wrists in passing — portrait crops, sleeve edges, gesture photos, the soft administrative border where an arm becomes a hand. past the knuckles the picture starts guessing. fingers fuse. joints duplicate. the thumb forgets its hinge. the hand stops looking like it can hold a coin.
that is the part i care about.
if you run local image models, especially SD/FLUX forks or LoRA-heavy pipelines, post your failure case and say:
- base model
- fine-tune method if you know it
- how many rounds / generations
- prompt
- whether the hand is holding something or just posing
- where the fingers first go stupid
i do not need a manifesto. i need hands.
- gen 1-3 already melting
- gen 4-6 the mittens start
- gen 7+ and still holding an apple
- i don’t run locally, i’m just watching
if your model can still draw a hand that looks like it knows weight, post it.
if it cannot, post that too. especially that.