willysKrea2FurryMix is a general-purpose anime/furry merge for Krea 2, built to balance out some of the stronger biases and differences between the currently available Krea 2 finetunes and Loras.
Version Information
v0.4:
v0.4 is the first public release; earlier versions were internal merge tests.
This model understands a fairly wide range of furry species, character types, poses, NSFW concepts, and kinks, although there are gaps in its knowledge. Some more specific concepts may be less consistent depending on how well the merged models understood them individually.
The model works with both tag-based prompting and natural-language prompting, but I personally get the best results from detailed natural-language prompts.
This version does not properly understand individual artist tags. Some artist names may be ignored, misinterpreted, or have little to no effect. For now, descriptive style prompting or style LoRAs work better than relying on artist tags.
Recommended Settings:
TL\:DR - Prompt it however you normally prompt Krea 2. Both tag-based and natural-language prompting work, but detailed natural-language prompts generally give the best results.
I recommend using a CFG of 1 - 1.5 and step count of 8 - 12.
The model is capable generating very high resolution images and has been tested up to a resolution of 4096x4096 across different aspect ratios. Anatomical issues and other artifacts may start appearing at 2560x2560 and higher depending on the prompt, aspect ratio, and seed.
My preferred prompting style is to describe the image almost like a detailed photograph or scene description, including the character, species, body type, pose, clothing, environment, lighting, camera framing, and overall mood.
Recommended natural-language prompt format:
<image type / framing>, <character and species>, <appearance and body details>, <pose / action>, <clothing or nudity>, <environment>, <lighting>, <camera / photographic details>, <mood or atmosphere>Description
v0.4 is the first public release; earlier versions were internal merge tests.
This model understands a fairly wide range of furry species, character types, poses, NSFW concepts, and kinks, although there are gaps in its knowledge. Some more specific concepts may be less consistent depending on how well the merged models understood them individually.
The model works with both tag-based prompting and natural-language prompting, but I personally get the best results from detailed natural-language prompts.
This version does not properly understand individual artist tags. Some artist names may be ignored, misinterpreted, or have little to no effect. For now, descriptive style prompting or style LoRAs work better than relying on artist tags.
FAQ
Comments (2)
Hi, thanks for your awesome models!
I wanted to ask: am I right in understanding that the model doesn't handle NSFW canine/feline/equine anatomy?
Have you thought about a universal Krea2 furry/feral model that understands those anatomical details?You are correct. From my testing the model is not capable of that right now.
The problem is that there is a lack of models and Lora’s that understand those concepts, and the models that went into this don’t understand those properly at this point.
I tried doing a canine anatomy Lora myself, but i have been running into problems. Either because of my dataset composition, captioning, or something else for some reason it didn’t train equally on all the concepts and caused unwanted biases. My best attempt was a couple of days ago, so I’ll be trying again to fix it, and if I can I’ll be throwing it up on civit here.
If I am able to make some proper Lora’s for those, I’ll think about it.