A hugely impressive development, DALL-E was just another step forward for the AI world. It garnered a fair bit of attention, generating news articles and social media discussions about what this technology meant for art and artists.

But interest soon died away, and beyond its users and the AI community, it was largely forgotten until April 2022 when DALL-E 2 was announced. This new and improved version vastly upped the quality of the images produced, as you can see in these comparison images for the prompt “a painting of a fox sitting in a field at sunrise in the style of Claude Monet.”

Again this sparked lots of discussions around the quality, the limitations of DALL-E 2, and its uses. There were also a slew of image examples shared, whether generated through the limited options on OpenAI’s site, or by people lucky enough to get access to the full model.

Based on our Consumer Research data, in the week of the DALL-E 2 launch, 7.6k people posted about it. A big help for this was people with access and Twitter accounts asking their followers to supply prompts. Unsurprisingly this was popular as users were itching to give the technology a go.

All in all, a decent launch for a relatively niche piece of technology, reaching a wide audience and clearly demonstrating its capabilities. We imagine OpenAI were happy with how it went.

But then something happened they didn’t expect: a tsunami of surrealist, bizarre, and sometimes unsettling images hit social media. And they weren’t coming from OpenAI’s technology, even though DALL-E was being mentioned alongside them.

To explain what happened, we need to go back to Jul 2021.

The masses join the party

On July 30 last year, Dall-E Mini went live on Hugging Face, an online community working on a range of AI projects. A product of one of their community week events and with no official links to OpenAI, essentially it was a recreation of DALL-E with far limited training. The main pull being that anyone could give it a go.

While a hugely impressive piece of work, it mostly went unnoticed on social media. It continued to pick up a few mentions every day here and there until the DALL-E 2 release. Interest slowly spread and by the end of May DALL-E Mini was referenced around 40 or so times a day.

Then, on June 4, this was posted to Tumblr by user lomleyo:

It received well over 200 shares that day, with interest on Twitter rising shortly after. Mentions of DALL-E Mini continued to rise, and by June 10 over 4.6k people were mentioning it specifically. Suddenly, you couldn’t open Twitter without seeing a 3×3 grid of strange and curious images, such as:

Others were far more bizarre and nightmarish, including the AI seemingly creating its own monster called ‘crungus’. In other words, the collective mind of the internet unleashed a wave of AI-generated fever dreams. Few could we include in this email, but feel free to search around Twitter. It won’t be long before you see something to give you cold sweats.

While this period also saw a decent amount of posts about the proper DALL-E, the main driver was the mini version. Thanks to it, there was far more interest generated than by OpenAI’s official DALL-E announcements.

But there was a PR problem for OpenAI. The fact the mini version used their technology’s name meant some people would think they had created it. Meanwhile people were sharing images generated by DALL-E Mini, and saying they were made using DALL-E.

Aside from the confusion this could cause, there were other issues. First, Mini’s images are much lower quality. Second, Mini was being used to create some particularly edgy and questionable images. While not a full-on PR disaster, it risked people associated OpenAI’s technology with memes, shock-effect jokes, and poor quality.

Luckily there was an amicable resolution. OpenAI reached out and Mini’s creators renamed it to the pleasingly punny Craiyon. While this is a step forward, it came quite late in the game.

All in all, it’s a clear example of the double-edged sword of virality. DALL-E is most certainly far more well known thanks to the viral mini version, but how many people know what the true technology is actually capable of?

OpenAI certainly isn’t being passive though. It’ll take some work, but it will be possible to fix things. We reckon designing the cover for Cosmopolitan is a perfect start.