Hello everyone,
Today I’m uploading a requested model from Bleach: Retsu Unohana — specifically her version that fights Zaraki in the TYBW arc, in a Tite Kubo manga style.
Note: You should always include the “Bleach” style tag in your prompt to make her look like she does in the manga.
Note 2: Ignore the version number in filename. I made two other versions before this one:
Version 1.0 was trained with 48/24 dim/alpha, but that turned out to be too much, causing some hallucinations in her haori.
Then I tried a higher learning rate, but the character started to overfit very quickly.
The final version was trained with 32/16 dim/alpha (the default for most creators) and a smaller learning rate.
I chose Epoch 10 as the best checkpoint of the final version.
Prompt: look at "Trigger Words" on the right side 👉
Changelog: look at "About this Version" on the right side 👉
My Settings (Forge Classic Neo)*:
*note, the example images only use high-res fix.
Checkpoint: waiNSFWIllustrious_v160 170
Sampler: Euler a, CFG Scale 7 6
Scheduler: Automatic Karras
Samplings Steps: 30
Hires.fix: R-ERSGAN 4x+Anime6B, 20 Hires steps, Denoising Strength: 0.5 0.45
Extension:
Adetailer (highly recommended)
Description
Initial Release
FAQ
Comments (4)
I finally had time to try it out today — really great LoRA, props to the creator. The only thing is the eyes seem a bit off? They look kind of blurry, and even with restoration on, it doesn't help much. Is that down to the source data? Do you think there might be a way to improve this?
To be honest, I test my models with Hires. Fix most of the time. I'll do a few epochs without it to see if there's an epoch where the eyes remain stable. However, I think the dataset itself is the issue.
I could try creating a version 2.0 using images generated by version 1.0, but whenever I've done that in the past, the character tended to lose a small amount of her distinctive visual features.
@IcoNx937 I went back and took another look at the original images. Some of the eyes are indeed blurry, but there are also clear ones. I have a half-baked idea—what do you think?
1. For the images with blurry eyes, what if we remove the eye-related labels? My thinking is that if we leave the labels in, the model might focus too much on the blurry eyes during training.
2. For the images with clear eyes, maybe we need to be more careful with the labeling—like using specific trigger words such as “half_closed_eyes”?
3. And like you said, maybe generate a few images with clear eyes to replace the blurry ones?
Sorry if this is silly—I don’t know much about model training. If I’m wrong, just forget I said anything. Thanks for your reply!
@a864915836 To be honest, changing the captions wouldn't make much of a difference, since many of the manga panels have very abstract-looking eyes.
I uploaded the training data for you. If you take a look at it, you'll notice that almost every manga panel contains rather poor eye details. The problem is that if a large portion of the dataset looks like this, the model will learn those features because they have a strong influence on the training. At the same time, I can't simply remove those images, since I need the model to learn the outfit as well.
There are a few tagging improvements you could try:
Positive: detailed eyes
Negative: bad eyes, asymmetrical eyes, uneven eyes
I'm also going to compare epochs 9 and 10. I uploaded epoch 10 because it was the last epoch where the details were still changing significantly. The later epochs were discarded due to the onset of overfitting. However, after extensive testing, I actually found epoch 9 to be slightly more stable, so I'll probably upload it soon as well.
To fully fix the eye issue, I would either need to inpaint the eyes in the problematic training images or generate new images and use them to create a version 2.0 of the model. However, both approaches have drawbacks.
Improving the eyes through inpainting could result in odd-looking images because it would alter the original manga style. On the other hand, retraining with generated images would gradually lead to a loss of character details and overall fidelity.
Furthermore, even SDXL has its quality limits. Images generated without Hires. Fix and/or ADetailer typically won't achieve the same level of quality as images generated with those enhancements.