Lawlas's Yiffymix 2.0 (furry model)

Disclaimer: Despite the name "Yiffymix", this model is not associated with Yiffymix itself.

Note: All example images are raw results, upscaled without the use of inpainting.

▼About

Hello and welcome. This model is an upgrade from my previous model Lawlasmix. It was developed using AOM3 and various other models, so be sure to check those out.

It's capable of generating both SFW and NSFW furry artworks with consistent quality showing decent level of detail in hands, clothing, backgrounds and so on (for a SD 1.5 based model mind you). Compared to its predecessor, this model leans more towards the anime style and doesn't always require artist tags (it can still adapt styles if you do use them). Writing prompts for this model can be tricky for newcomers, so please read the tips carefully!

▼Tips

• Personally I use it with these settings:

  Batch count: 6

  Clip skip: 1 or 2 (for different styles)

  CFG: 6-8

  steps: 40-150

  Size: 512 x 704 or 512 x 780 .etc (then use hiresfix, or SD upscale in img2img )

  Sampler: DPM++ 2M Karras or Euler a

• Prompting

  The model can make amazing results with very simple prompts so long as you use it right. It uses both danbooru and e621 tags, like 1boy, 1girl, solo_focus. It's recommended to start the positive prompt with "(furry art, uploaded on e621:1.4)" and negative prompt with "(worst quality, low quality:1.4)". Since the merge has significant amont of anime models in it, you need to give more weight to furry related tags. Use tags like (anthro furry:1.6) in the positive prompts and (human:1.6) in order to get it to make furry art. The model performs rather well when you specify the color of the fur. For instance, (detailed red scales:1.5), (detailed blue fluffy fur:1.5). Feel free to check out the example images for their prompts to help you write your own.

• VAE

  Without using a VAE, the model’s outputs may suffer from color loss. Consider using anythingv3 VAE or orangemix VAE. The default option that comes with this model is orangemix VAE. If you prefer not to use any VAEs or encounter issues, use versions with a VAE baked into the model.

• Embedding

  Feel free to use other textual inversions, but I strongly recommend boring_e621. It's trained specifically on furry artworks and works great in this case. EasyNegative is another decent option.

Known problem(s):

1. Since the has AOM3 mixed in it, it's very common for it to generate results in which characters have "M" shaped hair. Consider using tags like (unique hairstyle, unique fringe, unique bangs:1.6) or using img2img to generate alternative images to help you mitigate it.

2. The model has reported to be temperamental so to speak. I suggest setting batch count to more than 2 so you have more results to choose from.

▼Credits:

Models used (as far as I remember):

1. AOM3

2. DivineEleganceMix

3. My previous model

Apologies for not keeping a detailed record of all models used. Without their amazing work, this model wouldn’t exist. Kudos to every creator on this site!