Custom model training for SD?

[mykeehu]

I wonder if Textual Inversion is not enough for me. I have trained in vain with different methods and images, but I have not been able to train it to what I wanted it to do. I have seen many different models trained on SD, but is it possible to train this on a local machine with webui (only from command line) or only on colab?
How should I get started? I read dreambooth description, but supposedly you need to train hundreds, thousands of images of models. How do I train a new model for SD?

[dill-shower]

https://github.com/Jack000/glid-3-xl-stable#trainingfine-tuning

This?

[mykeehu]

Thank you! Here it is already clear that I can't train the model at home, only on colab: "minimum 48gb vram to train". I only have 12 GB VRAM.

[ClashSAN]

here's another one: https://colab.research.google.com/drive/1vrh_MUSaAMaC5tsLWDxkFILKJ790Z4Bl?usp=sharing

[dill-shower]

Thank you! Here it is already clear that I can't train the model at home, only on colab: "minimum 48gb vram to train". I only have 12 GB VRAM.

You can use less VRAM if you set a smaller batch size. But such training is slower, much slower than TI, and it takes a lot of resources or time to fully train. You can read how Waifu-diffusion was trained (it seems that it was taught by this method): https://gist.github.com/harubaru/f727cedacae336d1f7877c4bbe2196e1#training-process

[mykeehu]

@ClashSAN and @dill-shower Thank you very much for the links and help! It looks like I'll have to dive deeper than I thought, as it will take more time. I'm not a programmer, just thinking of adding new tags/labels to the existing ones, but I may look at other models to see if TI might be enough for one of them. I'm enthusiastic, but I feel I'm moving slowly, which is why I'm open to more solutions, one of which would be model training. Thanks for all the help!

[kalkal11]

This one is my preferred to run locally, works on 12gb VRAM
https://github.com/bmaltais/kohya_ss

That repo is a simplified English verison of what's here -
https://note.com/kohya_ss/n/nee3ed1649fb6

(Google translate does a fairly good job if you'd like to read the contents of that second link)

[ClashSAN]

We think @mykeehu means finetune training of a model:

Finetune Training: 1000+ images, tagged, many new prompt words and new ideas recognized

Dreambooth: 5-50 images, not tagged, one "special keyword" and one idea (usually) recognized

These are the differences and they are regularly confused with one another.

[mykeehu]

Yes, I would probably add more words here rather than one, although I experimented with TI to see if I could condense what I wanted into one word and teach it separately, but I failed because I am missing words that I would refer to for teaching.

[mykeehu]

Although my graphics card is not big enough in VRAM, this trainer looks pretty good and fresh:
https://github.com/victorchall/EveryDream-trainer