illustriousXL/NoobAI Stabilizer

These LoRAs were trained because I want to reduce the chances of strange character poses and styles (stablize them), by learning more clean, natural, high-quality images, instead of only relying on tons of negative prompts.

It should also, in theory, slightly improve the overall output quality. (Becase it learned more "masterpiece" images ( classified by me :) ) without the "masterpiece" tag)

All cover demo images have no negative prompts, no quality tags prefix ("masterpiece", etc.).

Note that these LoRAs are "weak". (Low rank. No text encoder. Not over-trained. Comparing to those "strong" style/concept LoRAs). Which means: It won't affect/damage the original style of the model/other Loras, nor overwrite the prompts you give. Sometimes fewer prompts may give better and natural results.

Recommended strength: 0.4~0.8.

Trained on 1k hand-picked images that have:

• Excellent quality.

• Clear characters. (Clear character lines, no obstruction, no blur effect, etc.)

• Natural poses. Natural body proportions. (No close-up, exaggerated art, chibi, jojo pose, etc.)

• Complex backgrounds that are related to the current state of the characters (for example: bikini -> ocean/beach, school uniform -> classroom/school) (Most images). No meaningless abstract art backgrounds.

• Characters interacting with background objects in a logical way. (Some images)

Versions/Updates

Note: NoobAI is finetuned from illustriousXL. And some checkpoints merged NoobAI and illustriousXL checkpoints. If you don't know which one to start, use illustriousXL version.

illustriousXL v1.3:

v1.0 -> v1.3: Improved dataset tags. Improved LoRA arguments. Shoud be more stable and have less impact on image composition. Smaller tensor size (100Mb -> 30Mb).

NoobAI v0.10:

Test version. Trained on NoobAI epsilon v1.

illustriousXL 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.