My preferred settings:
Steps: 30-50
Resolution: 1024X1536
CFG: 4 - 5
As with most Illustrious models - close up framing will give much better character details and specifically yields significantly better eyes/face one shot outputs. I chose this mix because it showed strong performance with NSFW wildcards (random poses even at higher resolutions).
Hires Fix at 1.0X upscale (no upscale) 25 steps / 0.4 Denoise (using the same model) seems to yield better results even at higher base resolution.
SweetTooth V1 is the product of 2 merges:
First CopyCat+One Obsession
Second WAI+SweetPoison
The product of those two merges = SweetTooth
High-level recipe:
- UNet: 70% CopyCatObsession + 30% WaiSweetPolish
- CopyCatObsession UNet: 60% CopyCat + 40% One Obsession
- WaiSweetPolish UNet: 60% WAI-illustrious-SDXL v17 + 40% SweetPoison
- CLIP: CopyCat native (through all merges)
- VAE: CopyCat native (through all merges)
## Respect And Thanks
SweetTooth V1 exists because of these creators:
[DaSiWa Illustrious | Anime / SweetPoison by Darksidewalker](https://civarchive.com/models/1034922?modelVersionId=2350111)
[copycat-illustrious v7.0 by calculater]
(https://civarchive.com/models/823763?modelVersionId=1892440)
[One obsession v21_Anime by maxfeifei8]
(https://civarchive.com/models/1318945?modelVersionId=2823418)
[WAI-illustrious-SDXL v17.0 by WAI0731]
(https://civarchive.com/models/827184?modelVersionId=2883731)
As well as the upstream creators in the CopyCat lineage:
[Illustrious XL 2.0 by ONOMAAI]
(https://civarchive.com/models/1369089)
[MatureRitual丨熟メス儀式 Illustrious by EKLL]
(https://civarchive.com/models/994401)
[dupli-cat_flat(obsession/illustrious) by calculater]
(https://civarchive.com/models/1603883)
If you like what SweetTooth is doing, please support the parent creators.
They did the work - I just smashed stuff together until I thought it looked nice!
Description
FAQ
Comments (1)
This is my first attempt at a merge with limited understanding of what I'm doing. Mostly this is the result of testing around 50 different merge recipes and clip/vae/unet combos of models I like. The next version will be a more intentional mix with a goal of achieving better LoRA testing while maintaining a satisfying default style/color otherwise.
Clearly Stated:
I DID NOT MAKE THIS MODEL FROM SCRATCH!
I TOOK OTHER PEOPLE'S WORK AND SMASHED THEM TOGETHER. (some of them were just smashing together other people's work as well...I guess if you think about what diffusion models are...we're kinda all doing that...anyway).
That is all.
Thanks!



















