V0 Notes:

I had a good chunk of them backed up, and grabbed some of what I was missing off other sites that scraped the account and re-uploaded them. The rest I'm not too worried about right now, and I'll just remake in the future if I want them. I appreciate it though!

@worgman figured that but yeah I asked just in case

The training data comes from the CGs of the game and the standing pictures. I also will grab any promotional art I can find or the "bonus" items that come with pre-ordering and such and use it to supplement. A lot of the outfits in this series are a problem though with limited data. On those I do a few things. Cropping the image to take a standing picture from a single picture into ~6 by doing a cowboy shot, a upper body, lower body , portrait, and then usually a close-up of an area that has a lot of detail.

Then I upscale the crops using AI. After that I usually train my "v0" version. If I'm happy with it I'll release it as v1 or a draft. Otherwise I'll use it to polish up the images I have and to create new images to supplement the training data. This was the main trick I used to get the outfits to be reliable in my old LoRA. I haven't done that on the three LoRA I have re-released yet.

@worgman Thank you for the reply. It looks like it's going to take a lot of effort...

@japsduiqwwnfo Most of that applies to all-in-one LoRA. You can get decent/good results with just cropping. I try and go for as much flexibility in outfits and pose variation as possible. The bad thing if you just take the limited data for some clothing options and just go the crop route is you lose that flexibility. I would rather have the outfit fail to train than make the LoRA more rigid, but that's just how I like to make them.

It is definitely a lot of work though! I would usually spend about 6-8 hours on my old character LoRA. I would do test generations on one while editing the dataset on the next.