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.
Description
FAQ
Comments (4)
I managed to train up to 10k steps overnight. After testing steps 3k-10k, it still feels like 3k is the winner. 10k comes in at second place, everything between kinda sucked. 10k seems promising especially at 0.8 strength. I've kept 3k as the main version for now, feel free to try both.
For those who are lazy like me, I've included a super quick workflow that includes auto-prompting. Its attached in the images in the post with the cheetah holding the camera.
The reasoning model can be found here, just add it to your models/llm_gguf folder in the models dir on Comfyui:
https://huggingface.co/bartowski/huihui-ai_Huihui-gemma-3n-E4B-it-abliterated-GGUF/tree/main
Here's a potentially better instruction than the one I used (I screwed mine up a little):
https://www.reddit.com/r/StableDiffusion/comments/1p87xcd/zimage_prompt_enhancer/
The Searge_LLM_Node and the Searge_AdvOptionsNode are not able to be downloaded even if I try to fix the insulation in comfyui
https://www.reddit.com/r/StableDiffusion/comments/1qt8o4e/the_z_image_base_is_broken_its_useless_for/
So it looks like there's something weird with how Z-Image Base trains loras. The more pics you add, the lossier it gets. Explains a lot honestly. Unfortunately that means until a fix or something comes out, there's likely not much to be gained from trying to update this model by increasing the dataset etc.