Type | |
Stats | 626 |
Reviews | (44) |
Published | Sep 20, 2024 |
Base Model | |
Training | Steps: 60,000 Epochs: 4 |
Trigger Words | Ratatoskr Style Anthro furry flat colors |
Hash | AutoV2 06A55118A6 |
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 :
This model can do:
Photorealistic
Digital Art
Anime
Comic
Vector Graphics
Line art
Stickers
can write short text
etc
This is the Flux variant of Ratatoskr I fine-tuned on a merge of DemonFlux
Note: This on trained on a dataset at around 2500+ Pictures of the SDXL Version.
So the model is in some aspects still behind the sdxl model. and still in (WIP)
I recommend using the workflow with sdxl refining with the ratatoskr sdxl model
Note: the model is using an custom clip_l it's in the training and in the Ratatoskr Flux Workflow
You can test the workflow and model here:
Ratatoskr - Creature - Generator [Flux Version]
Changelog:
V1: Fine-tuned on 525 Pictures first version
V2:
so after the first " test version" with limited dataset i added another finetune with 2000 Pictures
so in total we now at 2500 pictures now so what has changed?
more consistent
some better details even without the sdxl refiner
better with lora use than V1
changed a bit the contrast level
better with nsfw
etc
to be honest this model was the most expensive i ever made so far over the time because i used external hardware again. so it will stay for now at this stage till i manage to use my own hardware.
Note: the model has a tendency for realistic and photo to correct that use the flat colors tag ;-)
V4
Model got a Huge Finetune trained at around 12K images overall, also added some own loras inside the model.
V4.10
Added an unreleased detail Lora i made into the model include a small amount of the flux hyper lora
Fixed some Color Degradation Issues caused possible by overtraining
fixed some precision issues
more stability
added some new specim and smaller details
model got a bit faster
This Model is in Early Access because ongoing tests (so this Version is still WIP)
More Details
more specim
also trained in some more art styles
better nsfw
V5.94
This model Version is Basically a new version based on V4.1 and my model Midgard Flux + new detail loras i made especially for this model.
Its unclear if this model is still a schnell model so i call it a hybrid Dev1. ;-)
the model is more precise
more details
overall more creatures
- but also a bit slower overall I can recommend this settings to start with:
Sampler: Deis
Scheduler: Beta
Steps: between 15-30 steps (25 recommended)
CFG: Between 2 and 4
There are 2 variants FP8 and GGUF (Q8)
BTW you can test the model out on tensor Ratatoskr - Creature - Generator [Flux Version] V2
An new workflow and a setting grid is in development.
Made by Freek22