Onetrainer a kohya ss alternative SDXL Lora Training on RTX 6GIG and UP

[@braintify]

I struggled with the installation for half a day, but in the end I won. This video helped me a lot with installing the environment - kzbin.info/www/bejne/Y3_Nf6xtlsuCh5I I hope this helps someone too.

[@streamtabulous]

so visual basics c++ i missed by looks, thanks will pin

[@streamtabulous]

Good news I'm working out the training I have a very ugly me so it's working. Update, lora looks great so have the configuration and settings needed. Unfortunately it's not using cudnn due to a issues with my backend not onetrainer so it's taking much longer than it should. I'll work that out next but this is fantastic news.

[@lechefski]

finally, someone making onetrainer tutorials! I'm looking forward to the next one! edit: I have the same 2 install error lines that you have- I'm using an RTX 4080 laptop (12GB) hope this helps anyone trying to get these two trainers to stop conflicting: my kohya issues disappeared when I reinstalled it but followed the github instructions properly this time now that I understand that I have 2 python version installed (1 for each trainer) "For systems with only more than one python release installed: .\setup-3.10.bat" now, if only I can get onetrainer to produce high quality 1.5 loras in ~25mins like I used to be able to with kohya...

[@streamtabulous]

@@lechefski yer i probably have to do a video on shuffling them the python order, that could be the cudnn issue as reading its fussy and only few python work with a few cuda toolkits, wish these AI tool had everything self contained so would not need system installation being every developer has different requirements

[@beragis3]

@@lechefski That's one headache I have had with the various pytorch appls, each one needs a different python verison, and that's the reason I use anaconda for easier switching between python versions.

[@MisterWealth]

What kind of iteration time should one expect ? I use a 3090 and 100 images at a time

[@streamtabulous]

be faster than my rtx3060 thats for sure, just be sure you have cuda tool kit installed. i deleted my screen shot of iters and have not made a lora for a bit, dad passed away. will do more videos just not ready atm. there is definitely more loras i will be making when i do.

[@2mnyshp]

When you tag, tag ONLY what the AI should ignore. Don't tag specific facial features of your face. Because you want yourself to be appearing consisten when generating images. Tag stuff that also changes like clothes, backgrounds, etc.

[@streamtabulous]

i like to remove my glasses and play with hair colour eyes etc found the tag helps with eye i have a odd green so the tag help point to my eye so i don't end up with base models eyes.

[@2mnyshp]

@@streamtabulous Ah yeah, then perfect of course!

[@eccentricworx]

Very easy to configre, very stable. Tried to train an embed, it was working well, except the embedding was always empty, only 8kbit big, no matter what tried. Will try a lora instead, let's see if this turns out.

[@streamtabulous]

git www.git-scm.com/downloads Python www.python.org/downloads/ Nvidia Toolkit - CUDA developer.nvidia.com/cuda-downloads Onetrainer github.com/Nerogar/OneTrainer

[@joneschunghk]

I have trained with this setting on my rtx-3060 8g vram. Its training speed is 1xx s/it (batch 4). Only takes 2 s/it (batch 1) on Kohya_ss. If I need to train LoRA to 5000 steps, it will take over 50 hours to train. I don’t think this speed can “alternate” Kohya_ss

[@PaulRoneClarke]

I'm showing my ignorance here. But after weeks of using Instant ID with absolutely superb results. Add a face, type in a prompt. 15 seconds later an image that is almost perfect of that face that works at almost work all head angles in different lighting etc. Then looking at what these model trainers require to achieve the same results. A degree in rocket science, a £1200 graphics card, hours of training on a large image set - and... after doing it, results that as SO much worse than Instant ID I'm wondering why the best results I get are from the simplest, cheapest and quickest method. (by a country mile) What is it that Instant ID does that all this faffing about with Loras fails to even get close to?

[@streamtabulous]

face swap is good i use it, but is limited regarding hear face type etc, though you can cheat and make sure images match head shape etc for best result. but lora training is more that faces its anything you want, full characters art styles anything

[@PaulRoneClarke]

@@streamtabulous Thanks for the thorough reply :) The face Loras I make are usually awful. I end up deleting them. I only have a 8 Gb GPU that might be part of the issue. will hopefully be upgrading to 16gigs soon. 24 gigs a 4090 - is way out of my price bracket as someone who only dabbles with AI.

[@marcusmeins1839]

Why is CUDA important? one trainer didn't have it before. now it is very slow when i train , slower than before without CUDA . The old version didn't have any of these new features. this is something new.

[@streamtabulous]

cuda is a direct link so programs can directly work with the GPU removing a bottle neck. it be more likely you might not have noticed prior and it just seems the way you see. the cuda tool kit speeds up all local programs that use AI, python etc where the gpu is used.

[@AceOnlineMath]

you should try stability matrix to install it, and part of your problem is over tagging your data, try just putting a random key word and man,

[@AceOnlineMath]

the random key word should be something that you wouldent expect the model to understand like _me123 or something and then use that in your prompt eg '_me123, 1man standing in a field'

[@beragis3]

A number like 3e-06 is not a code, but a representation of a number in scientific notation. For example 3e-4 is 0.0003, 3e-6 is 0.000003.

[@streamtabulous]

thank you, so id imagine 0.000003 there would work also, always learning very much appreciated