Fgreat
UHello,
I'm using the configuration you have been using for 48 gb ram. I rented an H100 GPU with 20 image dataset but the images generated after training does not even match with my training data. Where do you think I might be going wrong? all the parameters are loaded from 15 image dataset config file. Can you please advise?
I'm using the configuration you have been using for 48 gb ram. I rented an H100 GPU with 20 image dataset but the images generated after training does not even match with my training data. Where do you think I might be going wrong? all the parameters are loaded from 15 image dataset config file. Can you please advise?
Fwhich config you used which zip file
Fplease more info
UI used 48GB_GPU_28200MB_6.3_second_it_Tier_1.json
Fok it is a dreambooth config. did you load into dreambooth tab?
Fproperly did set training images folder?
Falso we have an example dataset
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Uyes I'm loading them in the dream booth tab. I also am loading the training images properly but just to be sure is there a way for me to know if the training images are being loaded properly or not?
UI can see the folder being created after clicking prepare training data
Fok this is accurate
Fyou can look at the cmd logs
Fto see
Fif you copy paste here i can understand
UI get this when I press prepare training data
/kaggle/working/kohya_ss/training_data/img/1_ohwx
man...
00:37:48-667704 INFO Regularization images directory is missing... not
copying regularisation images...
00:37:48-670635 INFO Done creating kohya_ss training folder structure at
and when I start training, you can see them in the message.txt file.
Thanks a lot for your help.
/kaggle/working/kohya_ss/training_data/img/1_ohwx
man...
00:37:48-667704 INFO Regularization images directory is missing... not
copying regularisation images...
00:37:48-670635 INFO Done creating kohya_ss training folder structure at
and when I start training, you can see them in the message.txt file.
Thanks a lot for your help.
message.txt546.13KB
UI tried other images including mine but didn't work, so I tried taking a few screenshots from the 15 image dataset of yours and at the 1000th step the sample image I get is not even close to the training images. I will try with Dwayne Johnson images now and give you an update.
Flast time i tested
Fkaggle was not able to train flux
F"/kaggle/working/kohya_ss/traini
ng_data/img/1_ohwx man"
ng_data/img/1_ohwx man"
Uoh? so what do I use? I'm using jupyter notebook through vast.ai
Fit says kaggle
Fis that accurate?
Fwell vast ai python version i see 3.12
Fvast ai template could be wrong i dont know
Fi test on runpod and massed compute
UI'm not sure either, because the interface I'm using is jupyter lab but the folders being created says kaggle.
Let me try using massed compute and I will update you with the results I get. Thanks for your advise.
Let me try using massed compute and I will update you with the results I get. Thanks for your advise.
Fsure you are welcome
Falso as a test you can use our example dataset
Fbut with flux
Fit is very easy to get maximum resemblance
UI'm using flux dev for finetuning. Where can I find your example dataset? I couldnt find them in the finetuing post. Do you mind sharing the link?
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Uoh sorry I thought you were talking about a different dataset. I'm going to use this dataset and try it out using massed compute now.
Uupdate -Dwayne johnson training works fine. I get good results around the 3000th step. I will try again for other characters to see if matches with the training data or not. Thanks for your help. I have one more question, is there a way to programatically select the best checkpoint instead of manually checking the checkpoints to see which one is the best?
UAnother question - In the dataset Dwayne johnson does not have hair but when I generate images a few images for Dwayne Johnson hair is being added, do you know why and how can we keep it real? I can add no hair in the negative prompt but is there a different way to keep it generic without hardcoding.
EHI! I have a question about lora-training and captioning. A friend of mine suggest that when you caption your object your not supposed to mention the object in the image at all, only everything around in like a semantic mask and where the object-tag works as the caption for the lora-object. On youtube-tutorials on the other hand everyone I've seen says that you should in detail concistently describe the object and only mention shortly the rest of the image.
What would you guys say is the case here?
What would you guys say is the case here?
Fyou need to evaluate yourself
Fyes it is about prompt
Fjust add bald :d
Fcaptioning is needed only when you do a general fine tuning
Fohwx man , ohwx class works best
Fthen after training in details give your prompt
Fdetailed captioning only reduces your trained subject resemblance
Fbecause even if you caption the elements you dont want to be learnt, it still leaks info of what you want to be learnt
EOk, so no captions only caption tag in kohya? Should I use a high or low network dim for 50 images that are very consistent, I'm training a bag
Fi recommend use my config and dont change
Fyou can use folder name as caption