
Roughly 7500 images of sourced fully captioned REAL human faces given to me by one of my friends, uncertain where they came from exactly. Seems like some are celebrity some are not, most are nameless to me because I don't know popular culture very well.
There is almost no body in the training images, so don't expect much but face and some overcooked clothing shifts. You can define clothing pretty easily due to Simulacrum and Flux's inherent clothing control. I dunno how well it'll play with other loras due to the LR, but it's an interesting fast outcome nonetheless.
So nothing lewd or NSFW from this, but it trained super fast for SimV4 using the new tested lora params
UNET_LR 0.0006
cosine with repeats
3 cycles
clip_l LR 0.000006
1600 steps
dims 16
alpha 32
It was less than 3 hours on 2 rented 4090s for 7500 pictures due to the training counts and LR I have set up, one epoch.
描述:
训练词语:
名称: v4_woman_shuffle.safetensors
大小 (KB): 156703
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success