Type | |
Stats | 7,407 |
Reviews | (332) |
Published | Feb 26, 2023 |
Base Model | |
Hash | AutoV2 8F54282B36 |
NEW MODEL RELESED
Disclaimer: despite having "yiffymix' in the name, the model itself has nothing to do with yiffymix.
Hello and welcome. This is my custom furry model mix based on yiffy-e18.
It's able to produce sfw/nsfw furry anthro artworks of different styles with consistant quality, while maintaining details on stuff like clothes, background, etc. with simpler prompts.
I personally use it with these settings:
CFG: 6-8
steps: 23-150
Size: 512 x 704 or 578 x 768 (then upscale it. Some of the example images here are upscaled.)
Sampler: DPM++ 2M Karras or Euler a
(Clip skip 1 is recommended. 2 also works but makes the style different)
For better results, it's recommended to use it with anything v4 vae file or you can use versions with baked-in VAE.
Starting the positive prompt with "uploaded on e621" and negative prompt with"(worst quality, low quality:1.4)" help too.
PS: Consider using bad-artist embedding or boring_e621 as well. it's optional to use bad-artist embedding and you can try other textual inversions as well. If you see this, don't be an idiot like me. the way to use bad artist embedding is by adding bad-artist instead of bad_artist and you'll see it doesn't nessesarily improve the quality of the result. Boring_e621 is overall more recommended.
Discussion
If anyone wants a pruned version that is fp32 without included training data (and not fp16), I uploaded a safetensors on Fileditch (renaming required, it randomized the name).
It's 4 GB instead of the almost 8 GB. fp32 usually gives some minor detail improvements, which is nice for people who are not limited by VRAM or RAM. Personally I prefer it even on my 4 GB card.
I felt motivated to upload after the hotlink to CivitAI within my Colab broke 😂 When you read this Lawlass, feel free to add the pruned version yourself! I did this with the help of the Toolkit extension for Automatic1111 👍
sha256 hash to verify yourself: 527b6d018134e44f6e16807af72f0627f5744cdf92dd636b32e49e752e60deb5
CLIP fix was not applied in the Toolkit extension, so that it's identical to the original 8 GB file
About to see if I can make anime furries with this model. (Usually use “inspired by Posuke Demizu and Aenami Alena”— anyone got better suggestions for modern anime style prompt keywords?)
Looks rather promising. Keep up the good work!
Mine are looking to blurry and filled with light, I'm using the 2 GB version and I forgot how to really use it, I could use some assistance.
Gallery
