Note (2/1/2025): Moving to new model pages. There are multiple base models and many old versions. Things are starting to get messed up. I plan to split this page to 3 pages. And maybe publish more different style LoRAs in the future for those models. In case you lost the link to this page, follow my account so you can get a notification when I publish the new pages and new versions (I guess this is how it works).

Note (2/11/2025): This is an open source LoRA, and its purpose is to benefit the open source community. Therefore, this LoRA is recommended to use with open source models, rather than closed-source models. To be clear, I'm not against monetizing checkpoints, I 100% agree with monetizing, that's how society works. I just don't like the way they do it. First released a half-trained checkpoint v0.1, under Fair AI Public License, which requires all LoRA must be open source. Then when the time comes, release the finetuned close-source checkpoint v1.0 to "harvest" the open source community. To be honest, 200IQ move. But I just don't like it, as a LoRA creator.

Stabilizer

Just a personal fun coding project. With lots of black magic.

I name the LoRA "stabilizer" because in the first few versions I just want to make NoobAI have a stable style (NoobAI style is very unstable because it has more "high concept but low quality" training dataset.).

But now it can do more than that.

It can improve character's face, eyes, hands, overall details, natural lighting, balanced brightness, enhanced contrast, anti oversaturation, ...many things... While not dramatically change the base model output.

Recommended strength: 0.4~0.8.

Version prefix:

• illus = illustriousXL v0.1 (also compatible with NoobAI both e/v-pred versions, maybe a better choice if you've stacked many illus LoRAs on NoobAI.)

• ani40 = Animagine XL 4.0

• nb_ep11 = NoobAI e-pred v1.1 (also compatible with v-pred)

About NoobAI v-pred: This LoRA is compatible with v-pred. It also can "fix" the oversaturation, just lower your CFG to 3~4 and do not use ancestral or SDE samplers. Rescale-CFG is not necessary and is recommended to be disabled.

All cover images are generated using only the base model in sd-webui with default settings, what you see is what you get, no inpaint fixes, no photoshop enhance, no negative prompts, no artist or style tags, no rescale cfg, ... etc.

If you want to improve this LoRA, comments and posts are welcomed.

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 an ~1k auxiliary "real world no-human" dataset with a little bit everything, landscape, animals, buildings, rooms, weather, lighting... to balance the main dataset.

Unlike concept/strong style LoRAs. This LoRA was not overtrained. It will not dramatically change the model output.

Update log

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 verison, effect is too weak, just ignore it

nbep11 v0.114

Big boost in contrast and CFG stability. Now you can get high contrast and "true" full color range at CFG 4~5 (v-pred is CFG 3~3.5, assuming you are using Euler and Normal Schedule) without oversaturation and deformed image.

This is done by feeding model with full range color images at training time. Which is the natural way to teach the model to generate full range color for every images. This is not the "noise offset" trick. "noise offset" simply forces the model to generate more extreme colors at every steps. It only teached model to generate extreme colors, but never tell the model what is full color range. So it does increase contrast but also result in low color range and oversaturation. Because "noise offset" is just a training parameter. It is very handy therefore it's recommended or even default enabled (e.g. Civitai online trainer) in most training tools. It's very common. If you stacked too many LoRAs or base models which were trained with "noise offset" enabled, you might not able to see the difference. The effect of this LoRA may be overlapped by the "offset".

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 improvments 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. Shoud 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.

This Lora is published on C.i.v.i.t.a.i. If you see this sentence on other websites, it means it is plagiarism.

Share merges using this model is prohibited. FYI, there are hidden trigger words.