A very versatile negative embedding that helps generate high quality images. Higher chances of producing normal looking hands in realistic images.
Best Practices:
Keep your token usage to below 75 for best effect. Most of the time, you don't need to add other quality specifiers to your prompts (e.g. worst quality, bad quality).
Usage:
Negative prompt: bad_pictures
If you notice any artifacts, you can try increasing or decreasing the weighting like this: (bad_pictures:1.1) or (bad_pictures:0.8).
Values between 0.8 and 1.2 seems to work best (clip size 2):
Tested on:
Description
Finetuned using 100 really bad images and pruned to 64 vectors.
FAQ
Comments (3)
What is "token usage"? Thx
A word converted to a number by the AI is a token, so a token limit of 75 means 75 words. The AI converts a word to a number.
If you use automatic1111 this plugin is very helpful. https://github.com/AUTOMATIC1111/stable-diffusion-webui-tokenizer
It helped me.
The word is a number my not be consistent between languages. Chinese words may have multiple numbers.
@Ispy23 Close enough, the prompt gets converted into vectors (tokens) by the model and the maximum vector a model can take is 75. Automatic webui just averages them out if you use more than 75 vectors, which dillutes the quality of the vectors.
The np_simple_negatives_v2 uses 64 tokens, which gives you enough room for image specific negatives like: nsfw, child, monochrome
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.