Just a personal fun coding project. Trained on a vast, normal and natural dataset. Can improve character's overall details, backgrounds, natural lighting, contrast. Reduce overfitting effects and bring creativity back......

The dataset only has high resolution images, and has ZERO AI image. So you can get texture and details beyond pixel level. FYI, cover images are not upscaled nor inpainted. If you think they are, then this LoRA did its job well.

Latest updates:

(4/30/2025): New SOTA utility LoRAs in Stabilizer collection.

SOTA? These two LoRAs are not from training. So they can do magical things normal LoRAs can't do. For example, ZERO side effect on style (not an exaggeration, it's really mathematically zero).

Contrast Controller: Now you can control the contrast like using a slider in your monitor. Unlike other trained "contrast enhancer", the effect of this LoRA is stable, linear, and has zero side effect on style.

Style Strength Controller: The mathematically true stabilizer you guys are asking for. Now you can reduce overfitting effects (bias, etc.) with zero side effect on style.

Effect test on Hassaku XL: Notice that model no longer has bias of high brightness and feels more natural at strength 0.25. (Strength < 0 amplifies the style. > 0 reduce the style).

.........

See more in the update log section.

.........

(3/2/2025): You can find the REAL me at TensorArt now. I've reclaimed all my models that were duped and faked by other accounts.

Stabilizer

Just a personal fun coding project. Trained on a vast, normal and natural dataset. Forcing on character's overall details, backgrounds, natural lighting, contrast. Can reduce overfitting effects and bring creativity back......

Cover images are the direct outputs from the vanilla (not finetuned) base model in a1111-sd-webui, simplest prompts, no inpaint fixes, even no negative prompt. They demonstrate the effect of the LoRA, not clickbait.

Share merges using this LoRA is prohibited. FYI, there are hidden trigger words to print invisible watermark. It works well even if the merge strength is 0.05. I coded the watermark and detector myself. I don't want to use it, but I can.

Remember to leave feedback in comment section. Don't write feedback in Civitai review system, it was poorly designed, literally nobody can see the review.

Have fun.

How to use

No trigger words needed.
You don't have to set the patch strength for text encoder (comfyui, etc.). This LoRA does not patch it.
Recommended -30% of your current CFG scale. This LoRA will increase contrast so you can lower CFG scale for more details.

Version prefix:

illus01 = Trained on illustriousXL v0.1
nbep11 = Trained on NoobAI e-pred v1.1 (compatible with v-pred)

For user that using "mix" models: There are many "mix/merged" models, which mixed illustriousXL AND NoobAI. Sounds cool, but those two models have a huge training gap of millions of images and steps. This just makes the "mix" model's compatibility really really bad. If you are using such model, try both versions.

Dataset

Only normal good looking things. No crazy art style. Not small (2k images, I won't brag it's big, there are many gigachads who like to finetune their models with millions of images).

Every image is hand-picked by me.

No AI images, we have enough of those. Training on AI images causes model to overfit and wipe out all high frequency details (textures) instantly. So everything looks like the same, flat and smooth.

There are 2 main dataset in the latest version:

a 2D/anime dataset with ~1k images. Character-focus (The characters take up most of the image. I also cropped and rotated many images if necessary.). Natural poses. Natural body proportions. No exaggerated art, chibi, jojo pose, etc.
a real world photographs dataset with ~1k images. Contains nature, indoors, animals, buildings...many things, except humans.
- Why real world images? Simple answer: Better background, lighting, pixel level details/textures. Advance answer: Avoid overfitting.
- You can read more about this sub dataset and why I added it, at here: Touching Grass. There is also a LoRA that was only trained on this dataset.

Older versions

New version == new stuffs and new attempt != better version for you base model.

You can check the "Update log" section to find old versions. It's ok to use different versions together just like mixing base models. As long as the sum of strengths does not > 1.

Update log

(4/25/2025): nbep11 v0.172.

Same new things in illus01 v1.93 ~ v1.121. Summary: New photographs dataset "Touching Grass". Better natural texture, background, lighting. Weaker character effects for better compatibility.
Better color accuracy and stability. (Comparing to nbep11 v0.160)

(4/17/2025): illus01 v1.121.

Rolled back to illustrious v0.1. illustrious v1.0 and newer versions were trained with AI images deliberately (maybe 30% of its dataset). Which is not ideal for LoRA training. I didn't notice until I read its paper.
Lower character style effect. Back to v1.23 level. Characters will have less details from this LoRA, but should have better compatibility. This is a trade-off.
Other things just same as below (v1.113).

(4/10/2025): illus11 v1.113.

Update: use this version only if you know your base model is based on Illustrious v1.1. Otherwise, use illus01 v1.121.

Trained on Illustrious v1.1.
New dataset "Touching Grass" added. Better natural texture, lighting and depth of field effect. Better background structural stability. Less deformed background, like deformed rooms, buildings.
Full natural language captions from LLM.

(3/30/2025): illus01 v1.93.

v1.72 was trained too hard. So I reduced it overall strength. Should have better compatibility.

(3/22/2025): nbep11 v0.160.

Same stuffs in illus v1.72.

(3/15/2025): illus01 v1.72

Same new texture and lighting dataset as mentioned in ani40z v0.4 below. More natural lighting and natural textures.
Added a small ~100 images dataset for hand enhancement, focusing on hand(s) with different tasks, like holding a glass or cup or something.
Removed all "simple background" images from dataset. -200 images.
Switched training tool from kohya to onetrainer. Changed LoRA architecture to DoRA.

(3/4/2025) ani40z v0.4

Trained on Animagine XL 4.0 ani40zero.
Added ~1k dataset focusing on natural dynamic lighting and real world texture. More natural lighting and natural textures.

Above: Added more real world images. More natural texture and details.

Below: Has detail enhancement, but not much.

ani04 v0.1

Init version for Animagine XL 4.0. Mainly to fix Animagine 4.0 brightness issues. Better and higher contrast.

illus01 v1.23

nbep11 v0.138

Added some furry/non-human/other images to balance the dataset.

~~nbep11 v0.129~~

bad version, effect is too weak, just ignore it

nbep11 v0.114

Implemented "Full range colors". It will automatically balance the things towards "normal and good looking". Think of this as the "one-click photo auto enhance" button in most of photo editing tools. One downside of this optimization: It prevents high bias. For example, you want 95% of the image to be black, and 5% bright, instead of 50/50%
Added a little bit realistic data. More vivid details, lighting, less flat colors.

illus01 v1.7

nbep11 v0.96

More training images.
Then finetuned again on a small "wallpaper" dataset (Real wallpapers, the highest quality I could find. ~100 images). More improvements in details (noticeable in skin, hair) and contrast.

Above: Has a weak default style.

Below: No default style, all data is equally weighted. Pure anime dataset.

nbep11 v0.58

More images. Change the training parameters as close as to NoobAI base model.

illus01 v1.3

nbep11 v0.30

More images.

nbep11 v0.11: Trained on NoobAI epsilon pred v1.1.

Improved dataset tags. Improved LoRA structure and weight distribution. Should be more stable and have less impact on image composition.

illus01 v1.1

Trained on illustriousXL v0.1.

nbep10 v0.10

Trained on NoobAI epsilon pred v1.0.