Stabilizer illustriousXL

Stabilizer Just a personal fun coding project.

The name "stabilizer" is quite misleading. I named this LoRA "stabilizer" because in the first few versions I just want to make NoobAI have a stable style. But now I've added too many things. It can improve character's overall details, hands, backgrounds, natural lighting, brightness, contrast, anti oversaturation.....

The advantage of doing this all-in-one thing is that you don't have to stack tons of different LoRAs for each of those aspects. They were all trained together in this LoRA, and there is no conflict that might burn or blow up your output. ( all conflicts will be averaged out to zero during training)

Also, this LoRA was trained in a way that will only modify final details. It will not dramatically change the base model output. So it's compatible with basically all models.

Cover images might look mediocre or bad, because they are the direct outputs from the vanilla (not finetuned) base model in a1111-sd-webui with default settings and fixed seeds and simplest prompts, no pipeline inpaint fixes, even no negative prompt. They demonstrate the effect of the LoRA, not clickbait.

Recommended strength: 0.4~0.8.

Version prefix:

illus = illustriousXL v0.1 nb_ep11 = NoobAI e-pred v1.1 (compatible with v-pred) 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 know it works because I coded the watermark and detector myself. I don't want to use it, but I can.

Dataset

Tldr: Only normal and boring things. No crazy art style. Not small (3k images, I won't brag it's big, there are many gigachads who like to finetune their models with millions of images).

The main training dataset contains ~1k hand-picked images:

Character-focus. (The characters take up most of the image. I also cropped and rotated many images if necessary.) Clear characters. (Clear character lines, hands, faces, eyes, no obstruction, no blur effect, etc.) Natural poses. Natural body proportions. (No exaggerated art, chibi, jojo pose, etc.) High quality and full of details. ~ Wallpaper level. And some sub dataset, ~2k auxiliary "real world no-human" images with a little bit everything, landscape, animals, buildings, rooms, weather, lighting... to balance the main dataset, add more natural details and avoid overfitting.

Full range colors

Implemented after nbep11 v0.114.

It will automatically balance the things towards "normal and good looking". Think of this as the "one-click photo auto enhance" button in most photo editing tools.

Pros: Big improvements in brightness, contrast and CFG stability. Now you can get high contrast and full color range (almost 0~255) at CFG 4~5 (v-pred is CFG 3~3.5, assuming you are using Euler and Normal Schedule) without oversaturation and deformed image.

Cons: Normal = Not suitable for crazy bias effect. For example, you want 95% of the image is pure dark and only 5% is bright, instead of 50/50%.

Update log Note (3/4/2025): New version for Animagine4.0 will be published at this page. But i will still write update log at here. Because they all use the same dataset.

3/15/2025 illus v1.72

This is a quite big update.

Same new texture and lighting dataset as mentioned in ani40z v0.4 below. Brings RTX level natural lighting and super detailed texture to 2D characters. Tags that related to some kind of light sources, such as "lamp", "daylight", "soft lighting", may bring better lighting effect. Note: Those are not trained trigger words. You can try other things.

Other small changes:

Added a small ~100 images dataset for hand enhancement, focusing on hand(s) with different tasks, like holding a glass or cup or something. Can't say whether it's useful, as popular checkpoints already have similar enhancement. Anyways, the more, the better.

Hand v1.72 vs v1.23 on good old illus v0.1:

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 ani40zero.

The plan was to bring RTX natural lighting and super detailed texture to 2D characters.

So I prepared a ~1k dataset focusing on natural dynamic lighting and real world texture.

It was trained in a way that will only add small final details. So it won't blow up your output and create a realistic 3D character.

ani04 v0.1

Let's give Animagine 4.0 a try. Init version. In my test it fixed Animagine 4.0 brightness issues and brings much better contrast. Maybe too optimistic. Anyway this is a LoRA so you always can adjust the strength as you want. CFG 5 and LoRA strength 0.5 seems the best for most cases.

illus 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". See the section above.

Added a little bit realistic data. More vivid details, illumination, less flat colors.

illus v1.7

nbep11 v0.96

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

nbep11 v0.58

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

illus 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.

illus v1.1: Trained on illustriousXL v0.1.

nbep10 v0.10: Trained on NoobAI epsilon pred v1.0.