Hello everyone,

today I’m uploading Ururu Tsumugiya (Fullbringer Arc Version) from the 16th Bleach season.
This model wasn’t requested — it’s actually a byproduct of my Orihime Inoue (Fullbringer Arc Version) model.

This model was created using images from an unpublished version of her. Lately, I’ve been very unsatisfied with my results. I realized the output quality was only average at best, so I changed my workflow — even though it takes much longer, especially for characters with very little training data. Using high-quality images for training should improve output quality significantly.

Don’t get me wrong, I also put a lot of effort into version 1.0, but success is limited by the amount and quality of training data. Many requested anime characters only exist in 720p streams. Upscaling helps a little, but it also makes borders blurry, which results in mid-quality outputs.

So I’m starting this new workflow for LoRAs with fewer than 100 images and will see how it goes.

For Ururu I did have about 30 Images. This version got over 200 Images, for comparison.

Prompt: look at "Trigger Words" on the right side 👉

Changelog: look at "About this Version" on the right side 👉

quality tags I use:

oekaki, anime coloring, anime screencap, masterpiece, best quality, amazing quality

My Settings (forge)*:

*note, the example images only use high-res fix.

Checkpoint: waiNSFWIllustrious_v150

Sampler: Euler a

Scheduler: Automatic

Samplings Steps: 30 - at least use more than 15! I often see people generate with 8 steps. This it not enough xD (I mean, you should generate you pics the way you want, but the results should be better with more steps)

Hires.fix: R-ERSGAN 4x+Anime6B, 20 Hires steps, Denoising Strength: 0.3-0.5

Extensions (nice to have for little more Details):

Adetailer (highly recommended)

CFGRescale Integrated (forge)

SDXL Styles (anime)

CD Tuner

Detail Daemon

DynamicThresholding

FreeU

SelfAttentionGuidance