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 Furries but can do way more and can be used as an general use model :
This model can do:
Photorealistic
Digital Art
Anime
Comic
Vector Graphics
Line art
Stickers
can write short text
etc
The model can do SFW and NSFW
This is the Qwen Image Version of my Ratatoskr Model Series
for more information about Owen Image please use the link above.
This is an ALL IN ONE model so there is an Costum Clip and VAE Baked inside the model.
the model is tested to run even on 8GB Vram and 32gb ram (RTX 5060 mobile)
The model its trained on my own ratatoskr database and parts of the furry enhancer database . This its trained on my Ratatoskr Database and the Furry Enhancer Database so no real pictures or celeb people pictures are used for training.
for settings:
I use personal this setting:
Sampler: Euler
Scheduler: linear_quadratic
Steps between 25 and 35
CFG from 3-4
Simple Workflows are in the images but i also will provide some in the Training folder.
i also will make some new workflows public in the future (WIP)
Changelog:
V1.0 First version
V1.8 ( Edit Model)
It's alive ........
This Qwen Edit 2509 Model version is an "Frankenstein Model" it is made by replace entire layers of the original model with Ratatoskr Qwen Image V1 ones. layer by layer so there is only a skeleton left of the original one i also have added a small lora on it too. (6800 steps on 110 paired images)
this model perform better with anthro and furry images than the original one.
Like the name suggests this model is not designed to create images only edit ones.
can do sfw and nsfw images.
do to (maybe) destroyed layers this model runs very fast at 4 steps to edit an image
i highly recommend using the workflow that i have attached to this model
the settings are:
4 steps
euler
cfg1.3
beta57 or bong_tangend
Important this provided workflow use 2 custom comfy nodes made by me that only can be found attached in the training folder
V2.99 Qwen Image
For this version i used the same Database but added some new pictures that got generated with V1 and use more pictures of the ratatoskr database in general and changed the training method slightly with V1 because its now trained on my own system even that did take some time longer to save costs but the training took about 3 weeks....
I trained on the V1 Model this time for Natural Language and e621 so it now "understand" prompts better. I also tried to get rid of the female bias a bit.
overall i added around 90k steps in training
more detailed output
better prompt following
reduced the female bias
better with nsfw stuff in general
better colors
more realistic but also keeping its drawing abilities
for settings this model version didn't changes much from V1
Sampler: Euler
Scheduler: linear_quadratic or beta
Steps between 25 and 35
CFG from 3-4
I highly recommend using the advanced workflow that is in the training foöder to get rid of the qwen grids by using z-image for upscaling.
if you like to support my work please give a thumbs up i also interested what people can do with that model i also can highly recommend using it together with the furry enhancer
IMPORTANT:
Training such models is very time and cost intense so please help me by.
Please checking out my other models too and give like ;-) when downloading this model!
Disclaimer:
This disclaimer is not here without reasons. please have in mind this is like a virtual showroom and posting extreme stuff is like 💩 on my desk in real live 😉 and you are responsible for your own images that you generate keep that in mind.
made by Freek22