Inspired by messg's work, I bit the bullet and trained my own custom lora, tailored to the overhead slime images I like. That is, action shot, slime falling from above her, splatter, facial coverage.
This was a lot easier than I expected, and I got something decent with not a ton of input. 20 training images, simple captioning. It works pretty well already for the workflows I like, which involve image-to-image work and editing in between rounds.
Here are a few text-to-image examples from simple prompts. (Like: "A detailed, high-resolution photograph of a pretty woman wearing a ruffled white bra and jeans and sitting on a white platform with a skyblue background. She is splattered with green slime, which falls from above. She makes an O with her lips, surprised.") ) fwiw this is with lora strength .7 and guidance 2.6.
You need to keep going, experimenting etc. The entry for creating Loras is low enough to practice and improve. Someone posted on reddit, creating is a good Lora is easy, creating a great Lora takes time and practice but the efforts are worth it. At the minute, I'm deep into refining data sets and training parameters and it's so time consuming trying to get the model not to bleed concepts, enabling enough flexibility so you can pose in ways the image set doesn't contain etc. Balancing the correct strengths so you can use other Loras along side it. I think it cost me $35 of compute to run the last training for 21hrs. But... you can also get amazing results for $3-$4 on CivitAI too so the opportunity is there for everyone to play around with.
These are pretty good for your first time playing around with it! Have you uploaded this lora to Civitai online by any chance? I'd love to have a play but don't have a local system yet.
These are pretty good for your first time playing around with it! Have you uploaded this lora to Civitai online by any chance? I'd love to have a play but don't have a local system yet.
Nope, not on Civitai. I'm not really interested in owning something that public. I'm not using a local system, exactly, it's a Google Colab subscription with ComfyUI. Big learning curve there but you don't need a strong computer for it, and it's super flexible.
10/19/24, 5:30am: This post won't bump the thread to the top.