40
AI slop sure is somethin'.
(hexbear.net)
It will poison your algorithm to click this, but I appreciate that most of the comments think it's real: https://youtube.com/shorts/pLRGQNNkq9g
I was assuming these models train themselves on their previously produced data, so a bunch of dudes with a pregnancy fetish would've used the model and in turn "trained" it to something like this, getting further and further away from what it "should" look like due to it sampling AI generated stuff instead of actual photo and art references. Though I think I might be misunderstanding a problem these AI models have.
nah, they're "trained" in a specific (and quite long from what i understand) process using a curated dataset
if they trained themselves on their own output they would very quickly start shitting out absolute nonsense, like just a big smear of randomly colored pixels in the vague shape of what was prompted
I mean...* gestures broadly at most AI art * can you blame me for thinking that?
No, once a given checkpoint is made it's completely static. Even the problem of AI generated material being used as training data for subsequent checkpoints is overblown: uncurated and incorrectly tagged it's the same as mixing in bad data of any source, but you've got a bunch of hobbyists making LoRAs or finetuning checkpoints using hand-curated AI images that meet whatever criteria they set (like what I mentioned about the models can effectively synthesize new things by combining concepts with mixed success? Some people, particularly fetishists, basically try to make a LoRA or something that reinforces the cases of a model getting the concept right because they don't have a lot of separate art). I've also seen an example of someone actively trying to make a LoRA that was as messed up and full of common AI defects as possible to be used with a negative weight to drive the output away from those concepts.