Some of the "passable" generations from when I started using LoRAs, I was still basically guessing how to prompt properly and eventually got better. I also started to generate more realistic images shortly after this. The prompts used a lost to history (but they are bad enough to not be much use).
5 Comments
theblueredditrefugee@lemmy.dbzer0.com · 1 pts · 3y
Fascinating - do you know anything about training a LoRa? I have a niche fetish (macrophilia if curious) and I'd like to create a model that can generate images but I haven't been able to find much in the way of resources on this
I could probably dig up the web scraper I put together to harvest training data from a specific imageboard - getting text to go with it is gonna be a trick though since the images on that board don't have titles, just tags
LastChancePoleDance@lemmynsfw.com · 2 pts · 3y
Do you know how to use Automatic1111? if so, it can very handily preprocess your images and then https://github.com/Linaqruf/kohya-trainer can create the lora.
I think about a weeks playing and you would start to get decent output.
theblueredditrefugee@lemmy.dbzer0.com · 1 pts · 3y
Seems difficult but feasible, thanks!
SneakyFox@lemmynsfw.com · 2 pts · 3y
Sadly ive never been able to try, I can barely generate images as I only have 3gb of VRAM (and 4GB is the minimum). I did have a quick look and found this: civatai. I also took a look on the the Unstable Diffusion subreddit and found somebody who makes the type of content you are looking for. They don't use a LoRA but instead pose with action figures and then use stable diffusion to replace the figures with models, interesting approach which works pretty well but a lot more involved ofc
theblueredditrefugee@lemmy.dbzer0.com · 1 pts · 3y
Civatai seems awesome, I've gotta check it out, thanks!
Using action figures with stable diffusion sounds really difficult lol. I'm bad at posing stuff IRL, I once spent multiple hours trying to take a single nude photo to post on discord lol