@Dr. Furkan Gözükara do i have to run accelerate config for multi gpu training?
@Dr. Furkan Gözükara do i have to run accelerate config for multi gpu training?
Fit doesnt improve when using single layers methods sadly. but it improved when i was using 18 gb config
Fyes so wrong
Fswarmui way faster
Fi will look for this i think kohya can add
Falso model probably knows that person face very well already
Fjust replied this to kohya
Fcant wait to test it 
Fwe set it
Fi recorded the video
Fneed to edit and publish
Vhe's awake
VDoc, when you get a chance, and it's not urgent, but I think the windows version of Instant ID gradio is currently broken. Likely do to a change in a package preventing non-safetensor files from running. Error is: "An error occurred while trying to fetch thibaud/controlnet-openpose-sdxl-1.0: thibaud/controlnet-openpose-sdxl-1.0 does not appear to have a file named diffusion_pytorch_model.safetensors.
Defaulting to unsafe serialization. Pass
Defaulting to unsafe serialization. Pass
allow_pickle=False to raise an error instead."Vthibaud does not have a safetensors version, i think only bin
Dgood news, with TE lr 9e-4 prompting "myname photo" triggers my photo its actually learning the token associated with me, I'm using regularization images from the class person, and my face is not bleeding to other men.
Dis still training
Dto accelerate the tests I increased the unet lr to 1e-4
Dmy gol is find a configuration to train multiple people on the same lora avoiding class bleeding and bleeding between the trained subjects
Finstant id works but i think that error occurs
Fi need to check if it affects
Fnice tests
Tim confuse about something, is it better to train the lora on flux with 1024x picture or higher ? because in your Patreon it says "better resolution is better result, not lower that 1024"
Fhigher res yields better true
Fif you can afford to wait
Ftrain 1280 x 1280
TOh boi, i was thinking about training 4K pictures, but i suppose even with a 4090 it will take me days hahaha
FFby the way after 1536 i didnt test
Fi trained max 1536x1536
DI'm using my actual name "Diego" I didn't try with a unique token that are not present on T5, idk if this is good or bad, I suppose that weird tokens like ohwx are not present on T5 or it contains unique combinations of letters? that the token is present on both text encoders is good or bad for training a new concept?
DWhy are my lora files so big (2Gb)?
I trained them with Network Rank = 128, what value should I set to get ~200mb files?
I trained them with Network Rank = 128, what value should I set to get ~200mb files?
Dused the Rank_3_18246MB_Slow configuration
Fwell the thing is we use something rarely exists so that it can overrwite data
Fbut t5 thing is new in the workflow opposed to sdxl
Fbecause we save as float
Fsave as fp16
Falso we save rank 128 :d
DI know but now with T5 in the equation that is frozen, how reacts to tokens that are not present on it?
Dthat's the question???
Dreally don't know
Dthere is a way to know if a token is on T5 or not?
Feverything get tokenized
Fso it exists
Fbut it is single token or not unknown
Fi asked this to swarmui dev and he said we dont have t5 tokenizer :d
Flike clip
Ftime to work hard :/
F
Gis it c:/kohya_ai/kohya_ss correct? batch files in main c:/kohya_ai correct?
Gis it c:/kohya_ai/kohya_ss correct? batch files in main c:/kohya_ai correct? and then run the files creating the /kohya_ss folder right? I have this issue too, on 2 machines