>>3023705There official documentation of NoobAI quality tags at
https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5?not-for-all-audiences=trueThe relevant section is:
Quality Tags
For quality tags, we evaluated image popularity through the following process:
Data normalization based on various sources and ratings.
Application of time-based decay coefficients according to date recency.
Ranking of images within the entire dataset based on this processing.
Our ultimate goal is to ensure that quality tags effectively track user preferences in recent years.
Percentile Range Quality Tags
> 95th masterpiece> 85th, <= 95th best quality> 60th, <= 85th good quality> 30th, <= 60th normal quality<= 30th worst quality
Aesthetic Tags
Tag Description
very awa Top 5% of images in terms of aesthetic score by waifu-scorer
worst aesthetic All the bottom 5% of images in terms of aesthetic score by waifu-scorer and aesthetic-shadow-v2
... ...
As above, the training for "very awa" used the top 5% scoring images using waifu-scorer:
https://huggingface.co/Eugeoter/waifu-scorer-v3Pic related is the image that algorithm uses to demostrate its scoring system.
So "very awa" is supposed to make an image more like the high scoring images here. Conversely adding "worst aesthetic" to the negatives avoid images like the bottom scoring ones here.
This only applies to NoobAI and merges based on NoobAI, but since that's the most popular modern SDXL model for merges and such it applies to a lot of models.
