IDM vton is best
IDM vton is best
_Can we integrate that into this ?
Fyou can if you know programming :d
Fbut already have idm vton app why dont you check it?
Fi just completed a training for a client with CogVLM2 captions
Fand results looking stunning
Fstill comparing with raw prompt
Fi just did a new training for 40 images
Fafter 20k steps it started to became overtrained
Freg images used
Fraw captions still provides better likeliness
Fi feel like 200 epoch yielded best
F200 x 40 x 2 = 16000 steps
PWhat's the most recent koyha install/ guide?
There are so many Patreon posts that I don't use that I can no longer keep up with the ones I might use.
Is there some way to sort them that I am missing?
There are so many Patreon posts that I don't use that I can no longer keep up with the ones I might use.
Is there some way to sort them that I am missing?
Fwhere you want to install on windows?
Pyes windows 10
Dupdate, with 25 epochs 3 repetitions and the TE learning rate at the same learning rate as the unet, the bleeding I think is fixed, I'm no so sure yet because is undertrained, I will resume the training for 25 more ephocs but the results are promising
Fkohya gui has auto install bat file
Fitself
Fyou only need to have python 3.10
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözükara on Patreon
Fi plan to fine tune sd3 with huge dataset
Fnon-commercial use free
Findividual commercial 50$ and company 500$ . what you think
Fstarting with 100k then 500k then 1m and so on dataset size
Fbut first i need to find best hyper parameters :d
Fso many things to do so little time
NAny tip for train some style checkpoint for asian model? I have like around 8k of hq pics of asian people 50% woman / 50% male I already did the caption thing using LLava, so using ur config can be correctly?
M@Dr. Furkan Gözükara hello Dr. are you planning to make a video about sd3 training with kohya? can we use the same settings we used for sdxl training?
HCould work, would really just depend on how well the model performs.
Have you had any success with a trial training?
Have you had any success with a trial training?
Fuse our config and llava or cogvlm caption them
Fyes hopefully once i found good config and workflow
Fnot yet
Ffirst i need to research 
F2GB x/y/z plot
HDoes anybody know how to train a LoRA for inpainting base models?
Currently dealing with the error below with kohya_ss.
ValueError: Trying to set a tensor of shape torch.Size([320, 9, 3, 3]) in "weight" (which has shape torch.Size([320, 4, 3, 3])), this look incorrect.
Currently dealing with the error below with kohya_ss.
ValueError: Trying to set a tensor of shape torch.Size([320, 9, 3, 3]) in "weight" (which has shape torch.Size([320, 4, 3, 3])), this look incorrect.
Finpainting model training is different afaik
HWhat I've heard is that you need extra layers for training inpainting models.
Can it be done by changing the kohya_ss training script? If so, I'd be grateful if anyone could share the method / working version of the code.
(Or do I need to change the training data? I believe the data used for training inpainting models did require masking images.)
Can it be done by changing the kohya_ss training script? If so, I'd be grateful if anyone could share the method / working version of the code.
(Or do I need to change the training data? I believe the data used for training inpainting models did require masking images.)
Fi never tried sadly dont know @highjourney
HThank you for your response 
If anyone has made any progress in any way, I would be really grateful if you could share it
I'll also summarize any progress I make in writing and share it here!
If anyone has made any progress in any way, I would be really grateful if you could share it I'll also summarize any progress I make in writing and share it here!
Fsomething broken with gradio or runpod
Fwow with correct new performance parameters i got 2x speed on runpod with a1111 :d
Dupdate, my sd3 training is still suffering from bleeding, training the TE improves but is still there, I'm tinking to remove the class and just train the names without class and add the class in inference, or change the captions adding a comma "name, class" and the other option is to increase the TE lr doubling or tripling it up, I'm confused with sd3, it bleeds concepts a lot
DI'm training now without class, just the names, lets see...
Dthe base model dont have bleeding, for example it knows perfectly the diference between "Michael Jackson" and Michael Keaton" but I cant get the model to make the diference between "person1 man" and "person2 man" it bleeds, I confused...
Fgood ideas to test
Fi wonder how many images they had
Fthe power of SD comes from huge dataset :d
Falso we dont know if trainer scripts are working accurately
Dthanks, yes you right, I will keep trying with different settings, it bleeds a lot compared to xl it must be the hyperparameters or the scripts or a combination of both