****** Quality Tags

****** Quality Tags adds two quality tags to ****** Furry XL, replacing the existing best quality and worst quality. It is an SDXL successor to my Fluffyrock Quality Tags with an improved dataset and a bit of preference optimization.

Usage

Just two tags this time. You can play around with weighting and tag order to control the strength.

Positive: best quality
Negative: worst quality

Model Info

The goal of this LoRA is to improve the style of the ****** Furry XL model withou* ***ng any other style tags. It produces a colorful, cartoony style with thick lines but can do semi-realism if you bully it hard enough.

Improvements include:

• Training an aesthetic scoring model to curate a dataset based on my preferences.

• Balancing the dataset so it isn't dominated by a few prolific artists.

• Manual validation every epoch to stop training when the style stops improving. Just 2 epochs was enough (20k steps).

• Implement MaPO preference optimization to improve the style even after conventional training stagnated.

• Good compatibility with Compassmix and Bananastrike that preserves their natural language capability. The style is less cartoony but still aesthetically pleasing.

Limitations

• Complicated scenes have more anatomy/coherence problems. This is actually a problem with the base model, but I wasn't able to improve it even with preference optimization.

• I used a synthetic dataset for preference optimization (for each real image, I generate a "rejected" version with the same caption). This lets me build a dataset quickly, but doesn't capture the same nuances as real preference data.

• MaPO is finnicky and starts to produce stippling and gridlike artifacts when training goes too long, even though the style was still improving. You might get faint artifacts in rare cases, despite my best efforts.

Initial release. SFT for about 20k steps, then MaPO for 3k steps