Beta testing a new AI feature

Hey all!

I've been hard at work the past couple months on a ton of projects that should start seeing the light of day very soon. One such project uses AI to help generate artwork or inspiration pieces for your characters based on the physical descriptions you've given them. It works really well for some characters and not as well for others, so I'm inviting anyone who is interested to come join me (and others) beta testing the feature before its release. In order to get feedback quickly while I work just as fast, I only ask that you join our Discord so we can talk about what is working well, what's not working, and what could/should be improved next.

If you just want to wait for the full feature release, that's okay too! I don't have a date planned for when it will be public, but I'm really excited for it and hope that it's sometime soon, maybe in the next couple of weeks. Due to the costs of running the AI (and to make sure we don't overwhelm it), it'll probably be a premium-only feature at launch – however, you can use it unlimited and for free during this testing phase.

If anyone is interested, here's the invite link: https://discord.gg/bDE8g5YRzp and we're testing in the #notebook-basil channel. :)

See you there!

Good questions! There's a lot more to be written on this to lay out all the details (which I'll outline when the feature gets closer to release), but here's the short(er) version touching on some of the bigger aspects:

Right now, the AI model I'm using was trained off a modified version of Stable Diffusion 2.1's dataset. This 2.1 dataset is already a much more limited version of the original Stable Diffusion dataset; it takes the original dataset and filters out any images attributed to a specific artist or artist "style", any recognized images of celebrities, any images labeled with a copyright, and any NSFW/adult images. This is more of a "base" model meant to add your own images to, but I'm using it as-is right now (without adding any additional images) while I look around for an ethical collection of images that include consent or some form of opting in that should be able to improve the quality of images later. I'm definitely okay with a lower quality in the meantime if it means the training data is more ethically sourced.

I'm also not using any Notebook.ai data in the training or processing, including any uploaded images. I may add the ability for people to choose to opt-in any of their own images later to enable some additional features (for example, being able to upload a few images of your character and then generate more images of them), but AI data as a whole is a pretty complex subject right now and I'd want to make sure anyone interested would know all the facts before consenting to anything.

Finally, I'm also building the feature to let me swap out the model/data at any time so I can use tools like HaveIBeenTrained that add one more layer of filtering, which lets anyone search public image datasets for their own work and mark them for removal (or explicitly opt them in to stay). Obviously opting out after the fact isn't enough for an ethical dataset, but it's another layer of security on top of everything else to make sure any copyrighted art doesn't make it through the cracks.

There's still a lot to be done but I'm trying to take all the steps I can right now to make sure any AI feature I add doesn't come at the detriment to any humans. It's meant to be a helpful feature to help us humans be more creative and inspired, not replace us or make life any worse for any of us. :)

This is an awesome feature. I might be a good beta tester for this because I might get some weird results with my image pool. I already use DALL-E 2 to generate images for a lot of pages that I don't care about that much but still want something sharp and unique looking (eg. plants, locations, etc). It would be very welcome to just have an image generation tool built into the system!

Welcome aboard! The bigger variety of people testing, the better: both for the typical kinds of characters and also to get more diversity in characters as well. I'm especially looking for people with characters in underrepresented or minority groups to make sure the system works well for them, too. Your experience with DALL-E might also be nice for understanding how the underlying prompts are being created and what effects your changes to character pages might have. :)

In general news, the beta test is going well. There's been a lot of updates since this original post and we're moving fast. Here's a short list:

• Basil can draw your characters in 5 different styles (plus 5 more experimental ones) to better represent however you're envisioning them
• Basil can now also draw your locations in 3 styles (new today, so there's probably plenty to improve upon here!)
• You can now manually modify how important each field is to tell Basil what to focus on (for example, if your character's hair color is wrong, you can increase your "Hair Color" answer's importance to tweak results)
• You can now rate your generations (which are all private) so I can better figure out which styles are working well (or not), as well as which character traits tend to result in better/worse images (for example, making sure every race/sexual orientation has equally-good results).
• There's now a public stats page where I'm tracking use and some initial quality metrics (to make sure things are getting better day by day), but you can see it too. :) I'll add more to this page soon, too.

Lots more work to be done (especially regarding quality), but it's a good start with, I think, some promise. I'm very excited to see how useful it can be for quick worldbuilding visualizations, especially to give yourself (or an artist) some reference material to make something even better by hand. :)

Interestingly, the consistency checker and automatic recording of character/locations/etc while writing were some of the very first things I wanted to build through Notebook.ai way back when I was first roadmapping what all I'd want to do with the site. AI was nowhere near where it is today so I always thought of it as a "someday" thing, but I do think the tech is getting closer to making tools like that possible. The generators sound fun for my own worlds, too – they seem like they'd at the very least be nice little writing prompts, inspiration, and/or settings that fit into an established world that you can choose to explore more through your own writing.

There are definitely some big questions about where (if anywhere) AI fits or will fit into the writing process as a whole, though. I love writing and have for as long as I can remember, and I want to make sure I'm building tools that empower writers to create richer, more detailed, more consistent worlds rather than replacing those writers. There are also a lot of hot AI models right now that were trained on stolen or copyrighted content, and I want to make it clear that any AI I add to the site will only use ethically-sourced training data (for example, not scraping all of Ao3 and other writing sites!) – which inherently will make progress on these features a little slower than some other sites. Since my comments above, I've also started on some official guidelines for Notebook.ai's use of AI. I'm testing a lot of different AI uses and approaches that I think could be helpful in empowering a lot of writers and other creatives, and things like generators, consistency checkers, and chatbots to talk to your characters look promising as options for some people. I appreciate the suggestions and hope I'll have something more concrete to share soon. :)