In Norse mythology, Ratatoskr (Old Norse, generally considered to mean "drill-tooth" or "bore-tooth") is a squirrel who runs up and down the world tree Yggdrasil to carry messages between the eagle perched atop Yggdrasil and to the roots of the tree to Nidhogg.
This model is Fine-tuned for Animals/ Creatures and Furry but can do way more :
This model can do:
Photorealistic
Digital Art
Anime
Comic
Vector Graphics
Line art
Stickers
can write short text
etc
Model can do SFW and NSFW Pictures
This is the SD3.5 Large Model of my Ratatoskr Model Series
also available in SDXL and Flux
This is a WIP Beta Version of the model so it still have several issues that will be fixed over time!
Important:
Settings the model use really low CFG values between 0.8-2
with this Sampler: dpmpp_2S_Ancestral
Scheduler: SGM Uniform or Normal
20-30 Steps
update: My best results so far: at 1.2cfg and 20steps sgm uniform
Note it differs strongly and its very inconsistent at the moment.
This is a Wip model so if you find better settings please share them ;-)
Changelog:
V0.2 Wip Version First Release:
Trained at around 700 images of Ratatoskr Flux and Ratatoskr SDXL
known issues:
not consisted
missing legs or bad hands
sometimes ugly compared to flux or sdxl
slow as hell
etc.
Please check out my other models too and give like ;-)
Description
first wip Version
Settings the model use really low CFG values between 0.8-2
with this Sampler: dpmpp_2S_Ancestral
Scheduler: SGM Uniform or Normal
20-30 Steps
This is a Wip model so if you find better settings please share them ;-)
FAQ
Comments (3)
Curious.. Hardware requirements for SD 3.5?
Can I run it with a GTX 1080? Even on Forge? Although I think I know the answer, just looking at the size.
Nice to see some good SD3.5 large models.
I would like to ask for a quantized model for simplistic/remote workflows, or people with limited storage/ram.
@Foxdude mentioned ram limits.
I can run it with my RTX3060 12GB with forced fp8 (~1/2 the size) and offloading but i needed to rip out the unet and ran it with the default vae.
unet on gpu0, clip/vae on gpu1
Even after the butchery the model produced phenomenal results.
hi it's me again
i did "some" experimenting
almost 50% of my experimenting is just me throwing stuff at the wall out of desperation and most of it came to me in the shower/bath/toilet
useful data though
tl:dr
your model works with fp8 (not quantized but lobotomized)
use beta scheduler and ddim
10-20 steps
your cfg can be raised
use perturbed attention
let me know if the guide was confusing or just want a google drive link to the models instead of running some guy's code