Willy's Z-Image Anthro Lora

This anthropomorphic Lora for ZImageTurbo was trained on a dataset of 1150 hand selected images of various anthro furry characters that were generated using Willy's Noob Realism V-Pred v2 that were captioned with natural language using JoyCaption.

Not all species are treated equal, and there are some issues. I will be re-training the Lora at a later time when I can sacrifice 3 days of my personal computer.

Works with other styles as long as you prompt for them correctly.

The suggested weight when using this Lora is 0.5-1.0 and works best with longer descriptive prompts.

The "Experimental v0.1" version of this Lora has not been extensively tested, however seems to have issues with certain natural language prompts and may have unintended side effects with certain species or natural language prompt combinations.
Has not been tested in combination with other Loras.

Experimental v0.1:

Experimental v0.1 was the first training run of the Lora. It took about 3 days to generate the dataset, pick through and prune the dataset from around 10000 images to 1150 images, and an additional 2 days to tag all of the images using JoyCaption. Training the Lora itself on a single 4090 (which is the only hardware I currently have access to) took over 3 days.

Only after training this version of the Lora did I find out that JoyCaption incorrectly tagged several species of animal, leading to some unfortunate loss of knowledge in the final product although does not seem to cause issues with generating the species they were mislabeled as.

The species that were incorrectly tagged were sergals, kobolds, a couple avian species, and some species of rodent. (Although I had asked JoyCaption to tag all of these species using the name "anthro /species/" with the correct species listed for that section of the dataset.