Model card

Introduction：

Using a 5,000 dataset, I've attempted to blend the styles of 17 different artists, with each style roughly accounting for an equal share (300 each).

strongly encouraged to blend multiple style to explore new styles.

*I accidentally deleted Bubble-XLV1. I rewrote the model card, uploaded more examples, and renamed it to Bubble_styleXL. They are essentially the same model.

*The example images I uploaded were all results from using t2i, then upscaled 2x using R-ESRGAN 4x+ Anime6B in i2i.

Pormot format：

artist:_1, artist:_2,...
<1girl/1boy/1other/...>, <character>,
<general tags>, 
<quality tags>

artists list:

*avoid using an excessively high weight (>1.3) for any single artist tag.

artist: atdan

artist: misaka

artist: qizhu

artist: qing_wu

artist: misaka 12003-gou

artist: wd (1106592840)

artist: kamue

artist: nekoda (maoda)

artist: ru zhai

artist: swkl:d

artist: teddy (khanshin)

artist: vardan

artist: houkisei

artist: icecake

artist: nikishiko

artist: ririko (fhnngririko)

artist: shanguier

artist: tansuan (ensj3875)

Quality tags: best quality, high quality, normal quality, low quality, worst quality

Employed an aesthetic model to rate the artworks, then manually adjusted some scores according to personal preference, and subsequently categorized the quality into five tiers based on simple thresholds.

Resolution：

The model was trained for ARB 1024x1024 resolution, meaning you can utilize the standard SDXL resolutions such as 1024x1024, 832x1216, 1216x832, etc. Avoid using resolutions that are too small or excessively large.

*

Emphasis：

For sd-webui users, please use version>=1.8.0 and switch the emphasis mode to "No norm" to prevent potential NaN issues.

Author's Note:

• This is my first release of a checkpoint, and it may not be perfect.

• The model was trained with reference to the Kohaku XL series' training approach, utilizing the lokr algorithm from the Lycoris project, and merged back into PDXL V6 with a weight of 1, without further fusion with any other models.