Sign In

Uneven Eyes Expression [PonyXL]

275
1.8k
11.1k
122
Updated: Apr 28, 2024
conceptexpression
Verified:
SafeTensor
Type
LoRA
Stats
1,543
10,733
Reviews
Published
Apr 28, 2024
Base Model
Pony
Training
Steps: 9,600
Epochs: 229
Usage Tips
Clip Skip: 1
Strength: 1
Trigger Words
uneven eyes
Hash
AutoV2
26DB34A491

LoRA to give characters an "uneven eyes" expression (aka. one eye is narrowed, the other is open wide). Can be combined with general expressions, such as angry, smile, disgusted and more. Also combines very well with my Gesugao Lora.

If you like the LoRA, please leave a reaction, review and/or tip. Enjoy!

Usage

  • Trigger word: uneven eyes

  • Recommended strength 0.8 - 1.0 (try scaling the strength of the trigger word before scaling the LoRA strength!)

  • Use Clip skip=1 or Clip skip=2 (other values may also work, use at own risk)

  • Slightly unstable, may be overpowered quickly by other LoRAs or "strong" style/expression tags

Example

See preview images (generated with pruned version).

Unpruned vs. Pruned vs. Lite

I offer three variations of the model, which only differ in the file size and quality. Pick the one that best suits your preferences. If you can not decide, simply pick the "pruned" version. Note that the quality is relative, even the lite version usually produces images which closely resemble the intended subject.

  • unpruned: The originally trained model. Largest file size, best quality. Pick if you want the best possible quality, at the cost of disk space.

  • pruned: Downsized model. Almost same quality as the unpruned version, with noticably smaller file size.

  • lite: Drastically downsized model. Smallest file size. Still has good quality, but noticeably worse than unpruned and pruned.

Notes

This LoRA was a lot harder to train than I expected. Mostly because it is hard to find images that show the expression well enough for the LoRA to pick it up correctly. I had to double the size of the training dataset using generated images, but the results are pretty good. Though the LoRA is not quite where I would like it to be, I'm happy with how it turned out.

Big thanks to @redtvpe for coming up with the idea and providing some initial training data!

Using training data

Feel free to contact me if you want access to the training data.