Choosing Negative Embeddings for Stable-Diffusion
Common Uses of “EasyNegative”
Hi, I’m kojirom.
I think most people use “EasyNegative” as a negative prompt for image generation using Stable Diffusion.
Comparison with Other Embeddings
However, is there no other choice?
To state my conclusion first, images generated with “EasyNegative” are of higher quality than those without it.
However, it was found that even better results can be obtained by using a different embedding.
Therefore, I am currently not using “EasyNegative”.
To see why I came to this conclusion, it would be quickest to look at the actual images.
Below is a comparison of images generated using various negative embeddings.
Experiment Results: Image Generation Comparison
Model and Embedding Selection
The embeddings used from left to right are:
・bad prompt version 2
・BadDream
・EasyNegative
・EasyNegative V2
・UnrealisticDream
・None
For a total of 5 different ones.
In this verification, we used the best 10 models that have been downloaded the most on CIVITAI so far.
I think you can see that there are many embeddings that can be used other than EasyNegative.
Next, let’s verify it with other popular models.
How did you find it? It’s surprising how much the results can change with just one embedding, isn’t it?
Image Comparison and Analysis
In order to create this article, I generated about 1000 images. When I picked my favorite images from each model, the embedding that appeared the most was actually not EasyNegative.
The results are as follows:
- EasyNegative V2 (3 images)
- BadDream (2 images)
- No embedding (1 image)
I even laughed at 3rd place myself.
Conclusion: The Importance of Embedding Selection
This verification revealed that unexpected good results can be obtained by paying attention to embeddings other than EasyNegative. I hope this article will be of some help in your exploration of AI image generation.