forum Visualize your world and its characters with ethical image generation
Started by @andrew health_and_safety flash_onAdmin

people_alt 103 followers

@andrew health_and_safety flash_onAdmin

Hey all!

You may have seen mentions of Basil around the site over the past couple of weeks during the soft launch to make sure everything worked properly, but this feature is now ready for everyone!

The feature, in short:

  • You can now generate images of your characters (and other pages) from what you've already written about your character – no setup, no copying and pasting specific things over, just choose a page and a style and Basil will do the rest.
  • If you like any of the generated images, you can save them directly to your character's existing gallery tab.
  • The feature is enabled for characters, locations, items, creatures, floras, foods, planets, landmarks, and towns.
  • Premium users can generate an unlimited number of images, plus all free users get 100 free images to generate as well.
  • No data (text or images) was used in any way to train this AI.

I've already generated a ton of images for almost all of my characters and it's been really nice to put a face to a name, as someone who vividly visualizes characters one week and completely forgets what they look like the next. You can start generating images of your characters (and other pages) at this link, or by clicking "Image Generation" when viewing or editing any supported worldbuilding page.

Lastly, I also want to bring attention to Our Commitment to Ethical AI. You can read more about our baseline guidelines for any new AI-related features there, but I've also taken the following extra steps with Basil to align this feature closer to empowering creatives and feeling like a powerful tool to be used by them, not replacing them:

  • The underlying model is a limited version of the open source SD2.1 AI model, which takes the original Stable Diffusion model and removes copyrighted works, unlicensed images, and the ability to generate images in any particular artist style.
  • I've further fine-tuned the model on public domain photography and collections of images collected transparently and with an explicit opt-in for AI training, and use embeddings of public domain images to ensure the model's generations are guided by these images.
  • The site has also joined the Content Authenticity Initiative alongside Adobe, Nikon, and others, and I'll be following their guidance on how best to implement responsible AI.
  • All generated images include an invisible watermark that clearly denotes that they are AI-generated, and includes all of the settings and parameters used to generate them.
  • Because all images are transparently marked as AI-generated, they are ineligible for copyright through the U.S. Copyright Office. You are, however, still free to use them however you'd like, including as a part of other copyrightable works, which are still eligible for copyright even if they include AI art. More information here.
  • All image generations use high temperature and CFG settings, which results in a much wider variety of possible outputs. While this does mean you'll see low-quality images more often, it also means images can stray extremely far from anything seen in the training data. This should simultaneously make Basil not very well suited for producing specific images (you should commission artists for that!) but also make the feature more capable of producing all kinds of weird and unique ideas to use as reference material while you're worldbuilding.
  • I did all of my testing with a diverse cast of characters from all backgrounds, body types, ages, genders, and sexualities. There is bias in any data, but I've crushed a lot of it already and I plan to continue doing so.
  • Finally, I've also added a public stats dashboard to be transparent about how Basil is being used and what level of quality is being produced. There is also one specifically for characters that I also plan to add quality score breakdowns for race, gender, and other character traits.

You can read the full blog post (with a bunch of image examples) below:


I've been playing around with this feature for the last few days and I absolutely love it. However, I realized that I've heavily modified some of my page fields (eg. my "locations" fields are almost all different than the vanilla template). Because of this, when I go to use basil to generate location images, it doesn't know what fields to pull descriptive information from.

Is there a way to tag or mark custom fields to be included in basil's generation?

@andrew health_and_safety flash_onAdmin

Awesome, I'm glad it's helpful so far!

Right now there's no way to mark custom fields to be included, but that's a great idea. I'll see if I can get something like that into a release soon; I agree, that'd be very helpful!