Earlier this year, we got to indulge two of our favorite things: supporting the arts and technology to build cool things. We built the Audience Interactive System (AIS) to power the Spokane Symphony’s ‘Beethoven vs A.I.’ performance. The goal was to create a truly immersive experience, blending a classical art form with cutting-edge tech to draw the audience right into the show. Our solution involved a big screen that everyone could see, and the small screen in everyone’s pocket—audience members could scan a QR code to jump into interactive word-clouds, AI-powered trivia, and ranked Q&As. As makers, seeing it all come together was a blast.
The interactive features were a hit, but they immediately brought a new problem into focus: great experiences need great visuals. But as it turns out, existing image libraries are all problematic. They’re either expensive, difficult to search, or too content-poor to be useful (I’m looking at you, PowerPoint stock images). So, we figured we could solve this, too, by building a tool to fuel our creativity—and yours—with a curated, easily searchable library of the best AI-generated images we could make. Try AIS IMAGES.
Problem 1: Expensive
Reflecting the zeitgeist of today, AI solves many problems. Historically, to get images, one had to take photos, or be able to draw. Both take a lot of time, so naturally become expensive to commission, especially at scale. However, with the power of AI image generation models, what used to take days now can be done in about a minute. (Not to discredit artists, what they do is amazing, but if I just need a picture of a peony, which of these is AI generated?)
Creating an image from these AI models requires a prompt – some textual description of what the image should look like, such as “A vibrant etching of a blooming peony, showcasing intricate petals in hues of pink, orange, and gold, surrounded by delicate green leaves and softly etched background, creating an impression of a sunlit garden”, along some compute power. By hosting these image models on our own machines, we can create these images for the near-negligible cost of electricity. Writing these prompts effectively is an art in and of itself, but more on that later…
Read more about AI and the Arts here.
Problem 2: Difficult to search
Using words to find images is surprisingly difficult to do with code. Traditionally, a person would have to write words describing each image, and you’d have to hope that users searching for images use a similar vocabulary, a time-consuming and inherently limited approach. A more modern method is using image-tagging AI to write those text descriptions, but this still relies on getting exact words (I typed “large brown animal”, why doesn’t “bear” show up?). A more robust approach is to use different AI (a multimodal embedding model, specifically) to construct numerical representations of both the images and the words that users are searching for, such that items that are conceptually similar are assigned similar numbers. (This one’s a bit unintuitive. Imagine a picture of a dog is assigned the number 2, a picture of a house is assigned a number 25, the word “animal” is assigned the number 3, and we need to find which image is closest to the word “animal”, which would be the image of a dog at number 2 – the math is pretty easy for a computer to figure out).
As a result, now we can search for ideas. There’s our bear!
And as an added advantage, we can also compare images to images. Going back to our earlier example, we might have an image of a wolf assigned at number 1. If we were to use that image of a wolf as our input, the next closest image would be the dog at number 2.
We can even further subdivide images into chunks and search based on just the content of that chunk – as an example, if I were to click on just the moon within the picture with the wolf, I can search for just other moon-like pictures.
(Trying to get from one image to another just by clicking on relate subsections of images is a fun rabbit hole to dive into. See if you can get from a search for “gnome” to pictures of modern railroads. Here’s the link again, try it, it’s free!
If we use that previous bear image as our input, we’ve found my favorite image in our library, a very majestic other bear on a cliff.
Problem 3: Content-Poor
Arguably one of my biggest gripes with existing image libraries – they don’t have the images that I’m looking for. They lack the stylistic and subject matter diversity required to actually be applicable to most cases. It makes sense, how does one pre-emptively figure out what kinds of images there are? Yet again, AI can help. One of the appeals of AI image generation models is that they can create an image that matches almost any prompt you throw at them, but this is still limited by the imagination of the prompt writer. To solve this, we’ve employed more AI – this time a text generation model. Initially, we used it to systematically create 100 categories of images, and for each category created 10 subcategories, and for each subcategory wrote 10 prompts for various images. While this created a broad starting point, we’ve also spent a significant amount of time refining the meta-prompt that guides the creation of image prompts, as well as filling in additional missing categories by hand as we’ve found gaps in our library. And while this systematic approach gave us subject diversity, we also tackled the content gap with sheer volume—we’ve generated over 100,000 images and counting.
As a note, not all of the images are perfect – there are plenty of weird mistakes. But one of the biggest reasons our library is actually useful, and not just amorphous AI-generated slop, is that we as humans have gone through and curated a lot of it by hand. It’s a passion project that’s let us learn different ways to appreciate art, and explore what it even means to create art in an AI-centric world.
Everything we’ve built – AIS, the 100 k‑plus free image library, and whatever comes next – exists to spark creativity, stretch technology’s limits, and bring communities together. We’ll keep shipping tools and content that make it easier for you to dream up something awesome, then share it with the world. We can’t wait to see what you create, and we’re already tinkering on the next upgrade.