This is my first style lora for Z-Image Base so go easy with me on this one. Crazy timing too, I think Z-Image Base literally just became available for loras on Civitai lol. Looks like I'm the first to officially post one in the category. That'll obviously change once people update existing loras but a fun prize to have I guess.
Quick summary of the 2 versions. I recommend starting with V1, then experimenting with 10k if you feel like it:
V1.0: More consistent and has a cleaner look.
V1.0 - 10k: Trained on much higher step count. Higher risk/reward. I recommend running it at 0.8 strength.
This is a quick experiment I whipped together, please don't take it too seriously. It'll never be as good as a proper furry finetune checkpoint and I'm not even sure if it's better than using no lora, but its a fun experiment.
To save yourself a headache, I suggest passing a reference image to https://huggingface.co/spaces/fancyfeast/joy-caption-beta-one and generating a "very long" caption, as that captioning model was what I used to train the data. Prefix it with "boerbara" as the trigger word. I find the following works quite well:
"boerbara. This is a highly detailed, realistic digital painting of a muscular anthropomorphic...", followed by the joy-caption from the character species onwards.
There's not much info out there on best training practices for Z-Image base so I sort of winged it. Trained on 800 images taken from a mix of my own gens and datasets I've built up from illustrious models. You can see I've uploaded 2 versions. I recommend V1 as a start, but 10k does produce better pics imo when it locks it.
Sorry for the bland sample pics, natural prompting isn't my strong suit and I ran out of time for today to generate better stuff. The results are ok imo but a bit more cartoony than I'd like. That might be due to the prompt though.