A Flux Lora trained on SabuArt's later work, there are two versions, one trained on 100+ images and one trained on 16. I'm trying to learn how to do better, and have heard that few images can be better for FLUX
So with a little testing, the sketchy style comes through more clearly in the small dataset. but it's also less creative.
Description
This is the version with a 16 image dataset
Comments (3)
The images you trained both of these with still had signatures, text, URLs, etc. on them so the model assumes they're part of the training data. In non-Flux models you can attempt to suppress these in a negative prompt, but since by default Flux doesn't have negative prompts the model will generate those artifacts (even if you bothered to mention them in the captions). It's a pain to crop/retouch the source images but that's kind of why the artists put them there.
Yes, I thought that by putting those things into the prompt, then removing them when you generate an image it might suppress it, and maybe it does a little
Also more than a few of the large dataset images are cropped images, breaking down the multi-paneled pages. Which is maybe why it's less present there