Stable Diffusion 2.1: Better Human Generation, DreamStudio Gets Negative Prompting
Stable Diffusion 2.1 is released, bringing human generation improvements thanks to NSFW filter adjustments on training data. DreamStudio now supports negative prompting.
Created on December 7|Last edited on December 8
Comment
Stable Diffusion 2.1 has released today, staying true to the promise that model updates would be coming quicker than ever starting with Stable Diffusion 2.0's release. Dream Studio, the official online implementation of Stable Diffusion by Stability AI, also got an update to go along with this release.
What's changed in Stable Diffusion 2.1?
Stable Diffusion 2.1's release answers some criticisms that the release of 2.0 received, particularly on the generation of humans.
Stable Diffusion 2.0's training employed an overly aggressive NSFW filter to remove adult material from the training data. While that decision itself was controversial for many reasons, the result was that humans were less represented in the training data, resulting in poorer human generations overall (NSFW or otherwise). That, plus the wider focus on non-human subjects in the training data, had many users underwhelmed.
For Stable Diffusion 2.1, the NSFW filter was toned down so that it resulted in fewer false positives. The filter still exists, so compared to earlier Stable Diffusion versions, the more frisky end of human generation might still be limited, but in general, Stable Diffusion 2.1 can now generate humans better than 2.0 could.
What's new in DreamStudio?
Stable Diffusion 2.0 was found to work great with negative prompting, but DreamStudio's implementation of 2.0 did not support it at release. Now, with the release of Stable Diffusion 2.1, DreamStudio supports negative prompting.
Negative prompts can be added by entering | <negative prompt(s)>:-<some number>. Here's an example:
The left prompt was professional photo of cute corgi puppies, the right prompt was professional photo of cute corgi puppies | over saturation, illustration, black fur, noisy:-1.0.
Get Stable Diffusion 2.1 for yourself
Of course, Stable Diffusion 2.1 is open source and free to download for local use. Like 2.0, there is a 512x512 base model as well as a 768x768 optimized model. Head over to the github repository or Hugging Face model repository for more information and downloads.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.