Stabilizer for RouWei
I see tails, everywhere.
Note: Cover images are directly from a1111, at default 1MP resolution. No upscale, no plugins, no inpaint fixes.
What is this
RouWei is a large scale finetune from Illustrious v0.1 with a dataset of ~13M pictures (~4M with natural text captions).
Solved main problems with tags bleeding and biases, common for Illustrious, NoobAi and other checkpoints
Dataset cut-off - end of April 2025.
vpred + epslion.
Built-in color/brightness/contrast control: low brightness, high brightness, low saturation...
too many features...
Stabilizer is a tiny finetune LoRA with ~7k hand-picked images. Quick recap:
It can reduce overfitted noise because it was trained on clean, natural, high res real world images with strong logical and structural relationships.
It adds natural lighting, texture, details.
It doesn't have a default style. The dataset is very diverse. So no style shifting.
Dataset has full natural language captions. (<75 tokens)
It only patches U-Net. Does not patch text encoder.
How to use
Version prefix:
epv080b: (Recommend) Trained on RouWei epsilon v0.8.0 base (pretrained version in author's hf repo).
vpv080: (Test version) Trained on RouWei v-pred v0.8.0. High contrast, but less details.
Note:
epsilon version is much easier to use. It's just normal good old SDXL that has already been supported by all software. I personally prefer epsilon, usually has better details and textures.
v-pred version requires specific parameters and settings, see RouWei main page for more info.
RouWei is based on illustrious v0.1, so it supports most illustrious LoRAs.
Update logs
(8/3/2025) v080epb v0.5
Same as v0.3, but lower contrast for more stable and natural images.
(7/25/2025) v080epb v0.3
This is all most a full training on the "v0.8 epsilon base".
(7/24/2025) v080vp v0.1
Initial release. Let's see what we got.
The training strategy is very conservative. So it won't bring big changes to images. Most of the time it needs strength > 0.8.
Feedback is welcomed.