When T5 / clip-l testing commencing? Need info on what to use, LR to use, etc
When T5 / clip-l testing commencing? Need info on what to use, LR to use, etc
Mt5 very sensetive, kohya is implementing different learning rates soon for each
Mso save your time and energy and wait for it
JI'm using a beta of a friends trainer that has split LR - but honestly from my trials so far they work pretty well linked.
Regardless, this is why I'd like to see thorough systematic "research"
Regardless, this is why I'd like to see thorough systematic "research"
JWhat LRs are you using w/ Kohyas at the moment?
GYour on it now right? Lol this is it.
Mkohya told me 1/10th the base unet lr
Mso if you're doing 5e-05 for unet do 5e-06 for te
Myo what is that
Mim so cooked
Xmy flux model is now working inside of swarmui, i had to reintall swarmui and select "flex" in the step to install swarmui
X
PIs there currently any method to train lora for small details like that? I tried training but when I use it, there are almost no such small details or if there are, the text will not be the same. Anyone who has trained has any experience to share?
Dis it possible to train a model or a lora, to make this kind of comic images?
Fthis happens when comfyui torch is not up to date
Fthis doesnt tell the error
Fi may start today good reminding
Fnice one
Fnope i havent tried yet
Fcan be doable with a good dataset
PIf you have time, can you try it? Thank you very much
Fcurrently super busy but i plan in future product training
Hdid something update recently? entry not found in URL: https://huggingface.co/Kijai/flux-fp8/resolve/main/flux1-dev-fp8.safetensors. getting this error when installing FLUX.1-dev FP8 Version model on runpod, no issues yesterday
Fextremely stupid move
Fi dont know which one better but probably e4m3fn.safetensors
FThe flux1-dev-fp8 was just renamed to flux1-dev-fp8-e4m3fn to better reflect the format of the weights, nothing in the file has changed, no need to re-download that.
Fprobably i was using that link as well now i have to change all
Fi told him to reupload i hope he does
Hsounds good thank you
F2x upscale add those noise
Forg
Dit adds horizontal banding
Dwhat upscale method?
F
F2x of swarmui
Ddid you try upscaling with supir?
Fi will do that
Fsaving both for comparison later
F
Fman all images are just so amazing
F
XWILD!
NHi i need to hunderstand
i'm using kohya_ss gui
what i experiments it's that i cant let empty the area path repo.
i'm using kohya_ss gui
what i experiments it's that i cant let empty the area path repo.
NNo matter what I put, it creates a folder in the root of my repo.
What to do if i just want kohya to upload safetensors in my root repo ?
What to do if i just want kohya to upload safetensors in my root repo ?
Fi couldnt make this work actually
Fso we have a notebook to manually upload
Nyou upload manually !!!??? in one click i guess !!!???
�@Dr. Furkan Gözükara where can I learn about lora blocks, block training and layer training?
