It’s fair to say that 2022 has become the breakout year for AI-generated art. Anyone following social media has probably noticed the flood of artificial-intelligence-inspired memes appearing across newsfeeds, many of them created with DALL-E Mini, now called Craiyon, a publicly available image-generation tool that launched this year. More recently, three more apps — Midjourney, DALL-E 2, and Stable Diffusion — have also opened beta versions for public testing.
At first glance, all four tools seem to work in the same way: users type in a text prompt and the software returns a set of images produced by machine learning trained on millions of internet images. In practice, though, each platform comes with its own strengths and limitations. I spent time with all four to better understand how they work, where they fall short, and how to access them.
Craiyon, formerly DALL-E Mini, is likely the tool that introduced many people to the generative-image wave of 2022. Created in early 2021 by Boris Dayma for a Google competition, it was released to the public in April. It produces nine thumbnail images in about two minutes, which is reasonable for a free tool designed for everyday users rather than data scientists, and it requires no account or personal details. Since results are not stored, users need to save anything they want by hitting the screenshot button beneath the images.
The images Craiyon generates have a dreamy, sometimes nightmarish, fluid quality that would appeal to Weirdcore fans, but the model still struggles with the human figure, especially faces. Portraits and limbs often become strangely distorted, which can be amusing but also unsettling. Dayma and his team also note that the system is trained on unfiltered web images and may unintentionally produce harmful stereotypes. Craiyon was renamed from Dall-E Mini at the request of Dall-E 2 from OpenAI, due to confusion over whether the two projects were connected, and it remains free through advertising and donations. It is available at craiyon.com.
DALL-E 2, whose name blends Pixar’s “Wall-E” with Salvador Dalí, comes from OpenAI, the artificial intelligence research lab originally co-founded by Elon Musk, who is now just a donor. Beta access began on July 20, though users initially had to join a waitlist while OpenAI studied ethical use and necessary safeguards. As of September 20, the waitlist has been removed, and users can sign up with an email address and phone number. New accounts receive 50 free credits for the first month and 15 more each month, with additional credits sold in bundles of 115 starting at $15.
Although its interface resembles Craiyon’s, DALL-E 2 offers a more polished experience and generates four images per prompt rather than nine, with noticeably higher quality. It also stores the 50 most recent generations in “My Collection.” The platform blocks violent, hateful, and sexually inappropriate content, though its treatment of public figures can be inconsistent. DALL-E 2 is also much stronger than Craiyon when it comes to anatomy and faces, and its output is impressively versatile, ranging from photorealism to Blingee-style pixel art and art-historical styles such as Rembrandt or Hockney. Users can sign up at openai.com/dall-e-2.
Midjourney is probably DALL-E 2’s closest rival in terms of quality. Founded in 2021 by David Holz of Leap Motion, the independent AI research lab runs entirely through Discord rather than as a web app, so a Discord account is required. New users get 25 free image-generation passes, but those are limited to semi-public chatrooms for “newbies,” while paying users can message the bot privately. During my own testing, I even came across some shameless fetish art in the public chats.
In terms of output, Midjourney is on par with DALL-E 2 and in some ways more flexible, including the ability to choose aspect ratios for different image sizes. Holz’s goal was to create a tool that could “make pictures that look good,” and that emphasis shows in Midjourney’s photorealistic digital illustration style, which can feel reminiscent of Simon Stålenhag. The platform also lets users refresh for new results, upscale images, or generate variations from any of the four images in a set. To begin, log into Discord and visit midjourney.com to join the beta.
Stable Diffusion was the most complicated of the four to navigate. Released in August by StabilityAI, a British AI software company, it is built around open-source development and community participation, which can make the terminology on its website and FAQs difficult for newcomers. Its two main features are DreamStudio, a text-to-image generator, and “Diffuse the Rest,” an image-completion tool. Users can access “Diffuse the Rest” for free, sign up for Dream Studio with an account or through Google or Discord, and even download DreamStudio via Github if they have a gaming PC and enough storage.
“Diffuse the Rest” lets users upload their own drawings or visuals and have the AI complete them, while my own attempts at drawing produced some amusing mistakes. Using an image as the starting point for variations worked better and gave more relevant results. Stable Diffusion’s open-source model has been praised for encouraging inventive add-ons, but it also leaves room for misuse, especially because the software can be downloaded and run on a computer rather than only through the web. “Ultimately, it’s peoples’ responsibility as to whether they are ethical, moral, and legal in how they operate this technology,” StabilityAI CEO Emad Mostaque said. You can access Diffuse the Rest here, and create an account for Dream Studio at beta.dreamstudio.ai/dream.
Overall, these four tools have opened up huge possibilities for ideation and content creation for millions of people. Even when some are paywalled, they offer a fascinating way to visualize what exists in the mind’s eye without requiring formal artistic training or talent, which levels the playing field in both positive and negative ways. One thing, however, remains clear: none of them handle text well. For now, at least, the human hand still has the advantage there.
