illustriousXL/NoobAI StabilizerTD;LR:This LoRA can make the character pose and style more stable and natural.Recommended strength: 0.4~0.8. Theoretically, can go up to 2 if you want. (rank/alpha=2).No trigger words needed.Compatible with NoobAI v-pred model.I honestly don't know what to name this LoRA. "Stabilizer" is just the first word coming from my mind.IllustrativeXL and NoobAI are very creative models. So they often output strange/unwanted character poses and styles.Therefore I trained this LoRA with many high-quality character-focus images. And it did achieve my goal.So far the training dataset contains ~1k hand-picked images:Character-focus. (The characters take up most of the image.)Clear characters. (Clear character lines, no obstruction, no blur effect, etc.)Natural poses. Natural body proportions. (No exaggerated art, chibi, jojo pose, etc.)100% 2D high quality illustrations. No cosplay, realistic images, etc.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.It is not overtrained and the dataset is very diverse. Thus, it will not affect the style and creativity of the base model. This also means it may not have effect on model which is heavily finetuned (usually has a default stable style). You may need to push the strength up to 2.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, etc.Versions/UpdatesNoobAI:Note about v-pred: Latest version is compatible with NoobAI v-pred model, even it was trained on e-pred model. Although v-pred sounds like a completely different model, the main difference seems only at the U-net final OUT8 block. This LoRA won't patch that block.ep11 v0.96: More training images. Then finetuned again on a small wallpaper dataset (contains the highest quality images I could find, but only ~100 images for now). More improvments in details (noticeable in skin, hair) and contrast. Tweaked some parameters. Seems very compatible with NoobAI vpred 1.0, at least I'm not noticing effect downgrade in my tests.ep11 v0.58: More training images. Change the training parameters as close as to NoobAI base model. Fixed CFG shift issue in previous versions. Seems very compatible with NoobAI v-pred verison, even in dark env, tested on vPerd09R, reached my 9/10 expectation.ep11 v0.30: More training images.ep11 v0.11: Trained on NoobAI epsilon pred v1.1. Using new stuffs same as illus v1.3.ep10 v0.10: Trained on NoobAI epsilon pred v1.0.illustriousXL:v1.7: Same new stuffs as ep11 v0.96.v1.3: Improved dataset tags. Improved LoRA structure and weight distribution. Shoud be more stable and have less impact on image composition.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.
