illustriousXL/NoobAI Stabilizer

TD;LR:

This LoRA can make the character pose and style more stable and natural. Including face, eyes, hands, overall details, many things...

Recommended strength: 0.4~0.8. Theoretically, can go up to 2 if you want. (rank/alpha=2).

No trigger words needed.

However, highly recommend to delete all negetive prompts for better details. All cover images are generated without negetive prompt.

You may need to lower your CFG (-20%) after apply this LoRA.

Compatible with NoobAI v-pred model.

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

• Character-focus. (The characters take up most of the image.)

• 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.

• Yes. 100% hand-picked by me. I also cropped and rotated many images if necessary. Not just automatically selecting sub-datasets with high scores or anything like that.

Various methods are used during training to avoid affecting the creativity of the basic model. It won't dramatically change the output.

For demonstration, all cover images are just generated using the base model in sd-webui with default settings, no negative prompts, no artist or style tags, no other fixes,no rescale cfg, etc.

Versions/Updates

NoobAI:

Note about v-pred:

99% compatible with NoobAI v-pred model. Even if it was trained on e-pred version.

Will also adjust/affect v-pred model noise schedule a little bit. Less extrame conditions. Which in theory makes sampling scheduler more stable, no more blow-up, less messy noise and color shift. For example, Civitai online generator no longer generates messy noise: https://civitai.com/posts/11447947 https://civitai.com/posts/11439620.

Lower your CFG to 3.0~3.5. (Assuming you are using common sampliers, eular etc)

Updates:

• ep11 v0.129: Very experimental. Maybe an overkill. Added some furry/non-human images. Cut half the weights and half the training steps. Removed contrast enhance tricks in previous versions. So in general the effect of this version is much weaker than v0.114. Maybe too weak. Feel free to give feedback in the comments. Some xyz plots posts are much appreciated.

• ep11 v0.114: Added a little bit realistic. More vivid details, avoiding flat colors. So called 2.1D I guess? Changed noise schedule settings a little bit to make v-pred more stable, avoid extreme conditions. So some sampling schedulers won't blow up, which including Civitai online generator.

• ep11 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.

• ep11 v0.58: More training images. Change the training parameters as close as to NoobAI base model.

• ep11 v0.30: More training images.

• ep11 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.

• ep10 v0.10: Trained on NoobAI epsilon pred v1.0.

illustriousXL:

• v1.7: Same new stuffs as ep11 v0.96.

• v1.3: Same new stuffs as ep11 v0.11.

• v1.0: Init version. Trained on illustriousXL v0.1.

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