Type | |
Stats | 319 418 |
Reviews | (58) |
Published | Dec 15, 2023 |
Base Model | |
Training | Epochs: 10 |
Trigger Words | nursecomics |
Hash | AutoV2 E2A020AD64 |
I was curious what would happen if I took a bunch of raw comic book pages, did some very basic captioning, and turned it into a LoRA. To keep the look consistent, I used three issues from an old "medical romance" comic series (see if you can guess which one!). The results are, quite frankly, terrible. I'm not presenting this as a LoRA you should actually use, but rather to show what happens when you create one trained on the broader idea of a "comic book" page.
It's a failure, but at least it's an interesting and instructive failure. Obviously, everyone should consider this to be extremely experimental. You can trigger the LoRA with "nursecomics," or with most nurse-related phrases, as I used exactly none of the best practices one should use for captioning a complex LoRA.
That said, the layouts this generates are actually pretty interesting. I've found that giving the LoRA 5-10 more sampling steps to generate additional details helps.