Relying on gacha, there are often pictures of clothing worn on models
#1
by
HUG-NAN
- opened
Based on flux dev testing, the Lora weight is set to 1, simple+beta, and the image is generated in 25 steps. Images of people wearing clothing often appear, and even if the Lora weight is increased, there is a high probability of drawing the character image. Increasing the hyper8 steps of Lora results in a higher probability of generating a model image
Realized that also, the lora also helps with realism for some reason. I hypothesize that training on pictures of only clothes helped remove some DPO training