license: wtfpl
language:
- en
tags:
- stable-diffusion
- text-to-image
inference: true
Diffusion model trained on 500k image-tags pairs scraped from Furaffinity. Alpha still, expect more epochs, more training data and overall better results in the future.
"anthro, fox, male, general, by 100racs" Epoch 24, no inpainting
The tags contain the original FA tag list with tags appearing less than 40 times in total omitted, plus a tag corresponding the the general/mature/adult rating. If the artist also appears more than 40 times, an artist tag is added as well. The full list of tags and their number of occurences are available here. Training was done on TPUv3s using the LION optimizer.
Due to using data from Furaffinity, the model offers a wide variety of tags that aren't as popular as in other models for creating niche content. For example, the tag vore appears 25379 times in the dataset it was trained on as opposed to 4730 for FluffyRock. However, the preciseness of the tags on Furaffinity can also be left to be desired compared to e621, and as such it is recommended to merge it with FluffyRock if you want more control over your prompts.
How to use in Auto1111
Load up the safetensor file as well as the provided yaml file and put them in your model folder. Additionally, you are going to want to use CFG Rescale: https://github.com/Seshelle/CFG_Rescale_webui. 7.5 CFG and 0.7 Phi are recommended.
Discord and Help
furry diffusion discord, in the "custom-models" channel
Patreon and Donation
https://www.patreon.com/LodestoneRock
Credits
Original concept and data: amogus__enjoyer on Discord
Training: Lodestone Rock
Special thanks to patron and donators:
- Without Ordinary
- Zenko
- TheGreatSako
- Kadah
thanks to all of your generous donation!