training details
Could u please provide more details about traininig? Dataset size, LR, VRAM used, steps took, batch size etc.. :) + that training script would be also very nice! Thanks for great work.
Btw - are here any limitations on what this lora CN can do instead of "proper" CN usually found on like ZIT, Flux etc?
I don't remember the batch size and learning rate but for dataset you can use this https://huggingface.co/datasets/Patil/base-data and yeah lora controlnet won't give that good result which we get with controlnet architecture .
Thanks for the info and happy to see your dataset, if you grant me access :)
Anyways, do you remember the hardware, steps etc? Or roughly the costs so I know if it is real to play around with "normal money" or you need investors :D
We trained it on h100 on modal.com and I have accepted your data requests but don't forget to give credits . Have fun
Cool, thanks. So you preprocessed this datased with something like anydepth and trained on those pairs? How long it took on h100? Was it just one h100? How many epochs of this 60k dataset? If you dont mind those details.. would love to train something as well but already burnt sooo much money on many attempts mostly on z image turbo, so would like to have better success rate on krea 2 ;)
don't train it on whole 60k dataset .. take subset of it maybe 5 to 10 k .. it took 9 hours for training
what's your usecase i mean .. which controlnet are you trying to train ?