Fthey shouldn't be overcooked that soon
Fwell best way is generating images on the model you want
Fthen delete the outliers
Fbecause usually these models are extremely overtained and biased
Ffor example i deleted naked pictures it generates
Fthat one supports bf16
F57 images are for a person?
生sorry i not really know supports bf16 it means i cant test it?
F3060ti supports you can test
Fboth bf16 and fp16
生oh but when i trigger this on the new question happen
生
生
Fyes
Fi've been training with 50-60 images per person and it works fine
Fat least in the past
Fi think you need smaller model
F12b wont work on 12gb
生i use 7b model now it say this
生
Flooks like installation error
Fyou can try this
Factivate venv
Fpip uninstall torch
Fpip uninstall torchvision
Fpip uninstall torchaudio
Fand then this
Fpip3 install torch==1.13.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
生thx it can into the website now
生but when i question the ans will show error
生
Fwhat do you see on cmd
生
生it seem like my vram is too small?
Fif training uses DDIM or DEIS noise scheduling... why regularization images are generated using euler_a ?
Fshouldn't be generated using ddim?
Di have 1gb vram lol
Fand i still don't understand how to do regularization images!
Fsome of my images are like this:
F
Fwhy i would want to re-train my model with this? it makes no sense!
DLOL
DWTH IS THAT
DCURSED IMAGE
Dthats a doll mate
Fthat's a "photo of a woman"
Dwhat is that lipstickkkkkk 
Fthe class images stuff is very unknown!
Fye out of memory
Ftry smallest