VERSION 2.2 IS THE LAST UPDATE
after the release PDXL version, I figure this model work best with small token/ prompt. Inspect the image prompt for reference/ exmaple.
-Realism or semi-realism work best with lower CFG like 4
-Toon or Specific work best with mid-high CFG like 7-10. This may also saturate the color.
ABOUT
AI Beast Mix is a model merge to achieve semi-realism result for furry character with NSFW kink such as muscle or hyper. The model also can generate several style as well. Sometime the model generate human character which can be prevent by specify the furry species. If you're into hyper, this model capable support those kink and even work well with lot loras and lycoris. Prompt properly and this model will work wonders. prompt works well with e621 prompt tag.
Also, Hires.Fix is recommended to get better details and result.
Which one to Download?
Pruned fp16 - Suitable if you feel like to generate image only. May lack some data but still close enough generate image like full model. Good for those want to save disk space.
Full Model fp32 - Normal and Heavy model but contain all the data that you can use to merge with other model. Feel free to use this model to make your own mix.
VERSION 2.2
This would be the last update of the model. I can't find any way to improve the model anymore and its finally reach its peak. This version, I decided to tone down the detail to the reasonable amount and give proper lighting.
Description
I decide to push bit further on realism which is pretty hard. All I get is semi-realism which not really that bad honestly.
Improve details and color and reduce weird generation. Face still bad generate if subject too far
FAQ
Comments (8)
Odd... it only shows the 16 version
You mean fp16? The latest version 2.1 does not have fp32. Don't worry, using fp16 still works wonders.
Hi! This is the best model to create my tiger! Thanks for the new version!!
How were you able to teach the neural network to draw the connection from shin to paw in anthropomorphic characters so well and realistically? 0_о
I don't know how to explain the neural network stuff but I can share my method.
First I pick human model that I like and then I "Furryfied" the model. That's mean I have to pick furry model to merge with the human model. I usually go with 0.8 (%80) as the human model as primary (A) and furry be secondary (B). After that I generate several image to see the capability what It can generate. Doing this, I've add the furry data into that human model or otherwise but both will end up have both "RAW" data which why model size too big.
Then I do another set like above (Set A) with different human and furry model (Set B) and then merge them together (Set A + Set B). After that, I merge that new Set model with latest AI Beast Mix model to produce new one. That new result, if I feel it lack something, I'll make that model merge with lora that I can see fit. This one I use SuperMerger to have Model and Lora to be merge.
Regardless, more model you add into the mix, more data it have to be able to generate. However, certain model will overwrite/ replace data from the model that merge within. Therefore, you can see some furry model able to generate similar result with different variation and some model with generate completely different.
@Kotarou Thank you for such a big and detailed answer! I was pleased to receive such an answer and it is interesting to read ;)
Oh my freaking god this model is perfection
Which upscalers and VAE should be used with this model?
fortunately the model creator leaves the comfy workflow in all the images :)