For months, the “first” AI art program, DALL-E, has been hidden behind a beta wall that has limited access. It’s now open for everyone to try out, with a generous amount of credits to boot.
On Wednesday, OpenAI.com removed the waitlist to sign up for DALL-E, allowing anyone to join after registering for a free account. (The linked blog post includes a link to sign up.)
Each subscription adds 50 credits to your account, with each credit generating four 1024×1024 images from a single OpenAI server prompt. You will receive 15 new credits per month, although the credits do not accumulate. OpenAI has also placed content limits on the type of images you can generate, banning violence, sexual acts (including nudity), politicians and public figures. On the other hand, 1024×1024 image sizes are larger than other AI art generators and images render quickly.
DALL-E also supports outpainting, a relatively new AI art technique that lets you create variations on a scene in certain regions. For example, if you created a prompt that generated a scene where a fairy and a giant picnic on a cliff, your vision of what the scenery would look like could clash with what DALL-E generated. Outpainting lets you simply highlight or erase the backdrop with a virtual brush, and DALL-E will provide variations in that region.
DALL-E is it good? In certain scenarios, yes — and in others, you’ll find more success elsewhere. DALL-E seems a bit dumb, in the sense that computers are dumb: it favors explicit prompts and seems to take instructions literally. If you try a very generic prompt – “the castle of time” I’ve used before – you’ll probably get compositions that look like pictures of regular castles.
Likewise, “a starship enters a warp portal against the backdrop of a binary star, sci-fi, epic, cinematic lighting” gave me something that felt a little uninspired. “Promptcraft”, where users create detailed and specific prompts to create specific results, can help here. But even adding AI artists’ favorite inspiration Greg Rutkowski doesn’t do much for the final image.
Instead, DALL-E seems to work best with simple compositions. “A photo of a kraken emerging from the ocean under the Golden Gate Bridge” generated the image you see starting this story, which in my opinion is pretty good. “A robotic fruit bowl” produced a rather nice concept image, below. Take your cues from the automated interstitial images that DALL-E generates as it processes prompts, and you’ll have better luck.
Mark Hachman / IDG via DALL-E
Keep in mind, though, that this DALL-E AI art generator is only the first generation. In April, OpenAI moved to the more sophisticated DALL-E 2 – which is also restricted to beta access at the moment.
If you’re looking for more artistic compositions, stick with Midjourney, which also works on a credit system. For those with access to a gaming PC or GPU, however, we recommend that you try out Stable Diffusion on your own PC, which lets you experiment with as many comps as you have time. And remember, AI art isn’t just images; you can play virtual D&D with AI, generate artificial voices, and more in our AI art playbook.