After trying to torture my 3090, I bring you my first LORA. This magnum opus of having no idea about anything is brought to you through my pain and suffering of trying to get scylla without legs. In my hunt, I have amassed 61 images which then I tortured through my wholly expensive silicon. Single epoch, all pain, completely unskilled.
Trained at 768*768. Sweet spot seems to be somewhere around 0.8-0.9, however I'm not your mother and will not dictate where your weight should be. Use squidlegs as trigger.
More weight more squid, past 1.2, heavy artefacting. See plot for example. Works well with anime models, specifically Store Bought Gyoza (餃子Mix). The LORA might add more than just squid legs - generally expect hair and eyes to take a hit too. Artefact of the training and the dataset, beggars can't be choosers. Can proudly say, dataset has minimal splatoon content, and no mature content.
Preview images were purposefully generated with minimal prompting to see how the LORA works on its own.
Mileage with other models may vary.
Q: Why
A: No one else did and that one Chinese friend I have is more interested in making AI do clop than refined purposes like this.
Q: Why train at 768*768 when many lower GPU people cannot generate images at higher res?
A: I sincerely do not care. I want to try and get good at this. I want to push the envelope of what I can do and make with my GPU. If you can, splendid. If you can't, I don't care.
Q: Why eldritch tentacle horrors?
A: They make me smile