>>37116411>And is there any guide to train embedding with colabHaving done one myself I sort of wish I wrote down my process better or actually created a proper colab for it instead of just making some weird frankenstein of other ones. Maybe if I have the time in the next few days but realistically don't expect anything.
I started from this
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynband then removed or ignored the parts specifically about dreambooth, did some extra steps to install deepdanbooru, and just sorta stumbled along from there. Things to note:
- The setup puts the stable diffusion install, including the models, on your google drive. Hope you have enough space and don't mind google almost certainly snooping on you.
- While you can run it on free colab during low usage times of the day, it's sort of slow and none of the embeds I made on free worked very well. I eventually paid about $10 USD for 100 "compute units" (a vaguely defined measurement of time thats cost fluctuates depending on overall colab resource usage) and you should expect it'll take 10-20 to train a 10k embed.
- I'd recommend doing the initial setup, model upload, picture curating, etc before actually changing it to a GPU instance so that you don't go through your usage too fast. You're not charged for actually making use of the GPU, just for having an active runtime using it, so make sure you activate premium gpu from the runtime menu at the top when you need it and deactivate it (or close the session entirely) when you are done so you don't waste money
