Discourse has introduced AI-Generated Post Illustrations, allowing users to generate thumbnail images based on post content directly from the editor. Admins can enable the feature via the site setting ai helper illustrate post model and choose from providers like StableDiffusion, DallE 3, or self-hosted options Falco. While users like Tris20 and Jagster share generated images and critique AI’s tendency to hallucinate or misinterpret non-English text, others express enthusiasm for the creative potential. ThunderThighs requests granular control, suggesting the feature be limited to specific categories to avoid overuse in serious forums — a request sam confirms is technically feasible. Paul_King raises concerns about cost and paywalls, noting that while self-hosting is possible, GPU resources remain expensive. dsims asks if StableDiffusion can be used independently without an LLM to generate prompts — suggesting a “fake” LLM that passes raw text. NateDhaliwal inquires about mobile support. The feature is currently available on meta for testing, with related tools like the AI Bot offering alternative image generation methods. Users are encouraged to enable AI Helper to get started.
We’re excited to announce a creative new feature that’s rolling out for our Discourse users—AI-Generated Post Illustrations.
This new feature introduces the ability to add images to a post based on the post content.
How Does It Work?
You can trigger the AI-powered Illustrate Post feature directly from the post editor. Once your text is analyzed, you’ll be presented with a set of thumbnails tailored to visualize your post’s content. You can preview these images, select your favorite ones, and flawlessly incorporate them into your post.
To Site Admins
As always, we’ve made the feature compatible with multiple providers so you have the flexibility to pick what’s best for your community.
Currently, you can use:
StableDiffusion
DallE 3
Nothing
This will cover both people opting for paid services and those who prefer self-hosting their models.
How to enable / disable
Check the site setting ai helper illustrate post model.
A group of people celebrating the release of an ai feature on an asynchronous communication platform which can now illustrate written topics via prompts
The biggest issue comes from the very basics how AI is working: when it tries to build up a prompt from text itself its hallucination is bigger that just power to two. It just trying too much to get something useful to Dall E or what ever there is or will be listening. And for us non-english speakers rounds to translations to another translation changes too much context and idea itself.
The result is rarely something that should, could or would use. Well, that is my experience anyway.
But amount of used token is quite high for some reason. Higher that asked directly from chatbot using Dall E persona.
It is nice touch, though. But it suffers same weakness that AI helpers generally and what CDCK can’t fix: kind of alpha state of OpenAI and limitations of API based usage.
Would be cool if we could limit it to one category instead of the entire site. I think it would be fun in our “everything goes” category but I’m not sure I want it all over in our more serious product side. And our community will use it to exhaustion.
This would be great if there was no paywall involved in accessing an AI - as things stand, small community forums are unlikely to see any benefit
I note that Microsoft & Edge browser do provide ‘free’ access to GPT-4 AI image generation (up to 1024x1024px) via their Bing/Copilot feature. Is there any cunning way to harness this for Discourse?
Could there be a way to generate images with only a StableDiffusion key? (I’m not interested in the other helper features at the moment) I see currently you need a LLM configured to generate a more descriptive StableDiffusion prompt. So maybe a “fake” LLM that simply passes the post content as the prompt?
