/jp/ - Otaku Culture

>>43608335
>>43608363
The problem with fried LoRAs like this is that it usually means the composition as a whole has a ton of bias. It's less obvious since you are unable to see the dataset if you aren't the one who trained it, but if you could it would be very likely that an image exists in it with a very similar pose, composition, style, etc.
It's somewhat challenging to overcome currently because 1) most users have no idea what they're doing, tons of misinformation exists, and 2) the trainer most use does not yet support any sort of validation of loss to detect when overfitting begins. https://github.com/kohya-ss/sd-scripts/issues/193