The furry model that contains just about everything, and then some.
Trained with CLIP Skip 2, aspect buckets, and a token length limit of 900. Tags and potential aliases have been included as well.
Model base resolution is 640x640, but can also work at 512x512 and also be pushed out to 1024 or 1200 depending on your prompt.
Autocomplete extension CSV:
Discord for fun and shenanigans:
Description
First release
FAQ
Comments (14)
pickles please
the ai i prefer using doesn't seem to use safetensors
I've now included a converted version as .ckpt for any outdated or picky software about what models they can load.
@Jordach Error(s) in loading state_dict for UNet2DConditionModel:
Missing key(s) in state_dict: "up_blocks.0.upsamplers.0.conv.weight", "up_blocks.0.upsamplers.0.conv.bias", "up_blocks.1.upsamplers.0.conv.weight", "up_blocks.1.upsamplers.0.conv.bias", "up_blocks.2.upsamplers.0.conv.weight", "up_blocks.2.upsamplers.0.conv.bias".
Unexpected key(s) in state_dict: "up_blocks.0.attentions.2.conv.bias", "up_blocks.0.attentions.2.conv.weight".
blender gave me this error and i dont know what it means
Honestly the best model I have ever used
Not sure why, but I'm getting terrible quality results at 512x512. Its better at 1024x1024, but still not great. Nothing like the examples. Any ideas?
It's a 640 res base model - squares can reach 768x768 reliably but non square resolutions can easily breach 1024 on either axis. It prefers at least one axis larger than 640 for it to not do a stupid. Such is the curse of higher resolution models.
@Jordach Oh okay. I was using highres fix for 1024x1024 (which starts at 512x512) which explains why my high res still wasn't great. I tried at 640x1024 (portrait, no highresfix) and was getting much better results, but the results still seemed a little over-contrasted. I turned the CFG down to 3 and that seemed to help, but it seems like the contrast/edges are very sharp on this model if you don't specify any type of medium or style. But honestly, that can be a good thing. Thanks for the reply!
@Foxbite The unfortunate downside to CLIP skip 2 training is that artist styles do not contribute to the model's base understanding.
@Jordach I think I understand what you mean. If I don't specify the artist, my results look like they were drawn in mspaint by a 10-year-old. If I specify an artist ("by Zaush" for example), I get much MUCH better results, even at 512x512.
Definitely one of the best furry models out there. Like other models, though, it really struggles with feral horses
what vae should I use?