This was inspired by a contest going on over at Pixai and I decided "you know what? Why not make a model for this?", especially since the only other LoRAs that use this are for a Fortnite skin and one that seems to be for realistic models.
Might have scuffed the tagging since I added a space between "animal onesie". If people wish, I can redo the model and make it a one word tag instead, re-releasing it as a 1.1 model.
EDIT: I realized the scuffing was due to me associating "animalonesie" so I think I may need to do one more version where I name the animal version something else.
Description
Used the Step method and cooked the LoRA to about 3300 steps. Somehow it did not manage to overcook.
FAQ
Comments (15)
Edit: Nice catch on the animal part! Easy to overlook. Downloaded and following in anticipation of the other versions~
Unfortunately, renaming it to a different tag didn't seem to distinguish it enough
@MassBrainImpact Sometimes having both in the same picture (or no onsie at all even) as part of the training data helps to differentiate. Could help to use a unique tag if absolutely necessary, but often just expanding on specifics can assist. "blank_hood" "animal_themed, bear, bear_ears" "animal_themed, racoon, racoon_tail" "hood_up" "hood_down" etc. It helps the Lora in training get closer to focusing on what it should be learning in my experience~ (presuming you haven't already tried this. sometimes it's just bad luck😅)
@taiconan I'll give this a shot. Will reply again for results. I did have both tags in some images, so I'll try splitting the tags this time.
@taiconan I found the issue: There's only ten images that did not have animal ears or any of the sort on them. This is gonna be interesting.
@MassBrainImpact I know the feeling! Good luck, Boss~!
(And I've noticed, at least with object loras... it depends on the circumstances, but you may even want to remove some tags... like "1girl" and "hands" and such. As then it "draws the lora's attention" so your mileage may vary = ymmv)
@MassBrainImpact HEY! LOOKING GOOD! 💖
to control color of pajama I just need to add color before keyword? like "red onesie"?
Yep. Hadn't thought of doing that but it was a good thing you suggested it. Managed to try it out and it worked no problem.
Try going to the end of your prompt typing BREAK then put ( ex. color pajamas)
that makes a new set of tokens and it's going to use full power color on the pajamas, while not necessarily blending that color into all the other parts of the image
@d0000d Hmmm... Let me see if that works.
@MassBrainImpact play around with it, it creates a separate block of tokens and runs them together after the first block sequentially, allowing your main prompt to be run first on each step, then you kinda get the benefit of first order token power, but after the main prompt has run. I use it to try to keep eye/clothes/accent/background colors from all blending together. if they are still mixing, bracket them
Ex: (main prompt) sdgkrhg, uhgdj, ghdglhsd, glhrlgh <loraCue> BREAK [blue onesie] BREAK [pink bow on head]
What is the base model you used for the showcase pictures?
Anything V5 (the vanilla version, not the ink one that's the most popular)
@MassBrainImpact Thank you!