:D
0.2 & 0.1 were trained with WD14 tags (danbooru), this final one uses the correct ones (e621)
Description
The trigger word isn't needed but it's there because... reasons
FAQ
Comments (7)
Sir, you are knocking it out of the park with your LoRAs, know that your contributions are very appreciated. Did you follow or can recommend a specific guide for training them?
https://files.catbox.moe/5c0dvj.ipynb
https://files.catbox.moe/gqqg2b.ipynb
I use these two colabs, I prefer the second one because it offers more control over the dataset, is simpler and preserves the metadata but it has some problem where it may OOM during latent caching with a T4, but there isn't any secret to it, I just dump a bunch of images in a folder, auto-tag with WD 14 and train, I think the most relevant part is saving various checkpoints to pick the best, I'd also recommend you change "save_every_n_epochs" to "save_every_n_steps" (and setting it to 200) to avoid having to do math on the saving frequency as to not have to many or to little checkpoints being saved
Thank you very much, you're awesome
i miss using this artist on my prompts. syuro made the best anal gapes in 1.5
I'm seeing so many delicious femboys. I'm in love 😍😍😍😍
a very wonderful generator !!!
more like a butt generator :D (to keep it clean here lol), yes very good