Inspired by the popular Boring_e621 negative embedding https://civarchive.com/models/87781?modelVersionId=94126 , this is a negative LORA trained on thousands of images across years of data from different boorus with 0 favorites, negative scores, and/or bad tags like "low quality." Therefore putting it in the Negative Prompt tells the AI to avoid these things, which results in higher quality and more interesting images.
Pros:
-Generally increases quality which means more details, better depth with more shading and lighting effects, brighter colors and better contrast
Could be Pro or Con depending on what you want:
-Tends to generate more detailed backgrounds
-Tends towards a more detailed or even more realistic look
Cons:
-Many of the training images were low resolution sketches or MSPaint style doodles, if you are trying to generate sketches or doodle style work putting this in the negatives may be detrimental
-Many of the training images were black and white sketches or otherwise monochrome/grayscale, if you are trying to generate images without color putting this in the negatives may be detrimental
-Accidentally putting this in the positive prompt instead of the negative prompt reduces the quality of images
Test images was done using a weight of 1 with only this LORA in the negative prompt. You can adjust the weight of the LORA to change the impact, however in my testing the impact of different weights was minimal.
The first version uploaded "boring_SDXL_negative_LORA_SeaArtXL_v1" was trained on the Sea Art XL model (https://civarchive.com/models/391781/seaart-furry-xl-10) and is intended to be used with that model. The second version was trained for AutismMix SDXL (https://civarchive.com/models/288584?modelVersionId=324524). Please feel free to request if you want a version trained specifically for any other SD XL model.
Description
First model trained for Autism Mix, filename boring_SDXL_negative_LORA_AutismMix_v1
FAQ
Comments (6)
I've found that putting this at half strength in the positive prompt gives better results...for me anyway. More detailed and colorful, compared to the subtle change when used in negatives.
Thank you for leaving this comment! Since using this LORA in different ways can give different results I hope people will experiment with it like you did. It'd be great to see examples to show what you mean!
so is that LoRa for negatives? never used that
This LoRA is actually good, but you can't use it as described. a1111 webui does not allow LoRAs in the negative prompt. If you try to do so it does not even load the LoRA, it just interprets what you entered as text. So y'all have been just telling the AI not to be boring, autistic, negative or trigger happy with the confetti!
You can verify this by trying a large weight like 1000, which would cause the image to explode normally.
You can use this LoRA correctly by using it in the positive prompt, with a negative weight. -1 works, -1.6 is getting worse, and -2 is completely broken. I've only tested it with one prompt though, so YMMV.
Example to use, put in positive prompt:
<lora:boring_autismmixSDXL_autismmixConfetti_negative_LoRA:-1>
Surprisingly, putting just the name of this LoRA in the negative prompt as we have been doing seems to help quite a lot. I found it even has the magical property of fixing some completely broken images, to give good images. I guess we only had to tell the AI not to be boring, and I will continue doing this.
This is like the best LoRA ever.
No Token needed? As its based on boring_e621_v4 which has a token