NOTE, NEEDS THESE EMBEDDINGS FOR NEGATIVE PROMPTS
Use the above link for the Furtastic negative prompt embeddings, or images will not come out like in the showcase ones. Things will be very 2D without the embeddings.
Differences: Each version is like a variant. V1 is extremely realistic but rigid. V2 and V3 increasingly trade this in for more variety, better poses and expressions, and better NSFW acts. All are viable and have their own merits.
Introducing SnoutMix
A Mix of non-Furry and Furry models to create a great variety of anthro options, but bringing out more detail, still giving a lot of freedom to customise the human aspects, and having great backgrounds, with a focus on something more realistic. A big focus of this model is the transformation aspect, as it reacts well to E621 prompts and is perfect for inpainting those TF/Furry details. Works well with realistic character loras.
The gallery images are often inpainted, but you will get something very similar if copying their data directly. They are inpainted using the same model.
Usage Instructions:
This model performs best when still using the negative embeddings (see link in description). DO NOT use restore faces. Use the gallery for examples of how you can structure prompts that it agrees with.
V1-3 Like ClipSkip 2. Later ones prefer ClipSkip 1.
Recommended steps: 30-40.
Sampler: DPM++ 2M Karras
CFG Scale ~6 is what I use for reference images.
I don't use Hi-Res Fix, I use Controlnet Tiled instead.
Recommended smallest image size is around 640x640, seems very comfy with 640x768/640x888. SnoutMix 4 and beyond is great with higher native res (I use 768x960, but it can go higher)
Version 5 requires stronger anthro prompting as it's much better at the base human generations to allow for great anthro/TF inpainting.
VAE is not baked in, but vae-ft-mse-840000-ema-pruned.safetensors is highly recommended.
Description
A more complicated mix, more ideal for inpainting. It does realistic humans better, and then works best if treated as an anthro inpainting model. V4 is still the current king for overall anthro flexibility.
FAQ
Comments (13)
NEW VERSION LETS GOOOO
Ahh yes man made horrors beyond my comprehension
hi! link to the negatives is down. any way we can find those still?
Hi. I checked the link, it worked fine in the description. It just goes straight to downloading the zip with all of them in it for me.
v6 when?
So I just want to say. SnoutMix V4 is the best one, and if you switch your sampler to DPM3, it makes the model use e621 tags.
Rfneg, and Badhandv4 are my recommendations for negative embeddings to achieve desired results.
If you load all of the recommended embeddings before all of you negative tokens/tags, it sometimes will improve results.
V5 was definitely more built just-for-me in terms of how I use SnoutMix personally. I've definitely heard from most people that v4 was always the pinnacle so far.
How do I get the negative embedding to work?
Download them, put the four into your embeddings folder, and then from inside your WebUI you can put them in your negative (most WebUIs will have a separate embeddings page to add them on, like adding a lora) :)
Snoutmix pony?
I love this checkpoint and it's different versions. But it seems to have a hard time drawing two subjects in the same image. Is there a way to get it to do that?