Artificial Intelligence or AI is now actively being used to help create images or works of art. And art created by AI has become so good that such works are even winning prizes at art fairs. For instance, in September 2022, the New York Times had reported how a 39-year-old won the top prize at the Colorado State Fair’s fine art competition for an artwork that was created using artificial intelligence (AI).
Thanks to advances in artificial intelligence, such image generators are now highly accurate. You can feed them a very specific and odd description, say a ‘man riding a pink horse in the snow with Taj Mahal in the background’ and get novel outputs. This is made possible by training the AI on a large swathe of images with associated meanings, picked up from all over the internet. While some of the AI models like DALL-E Mini are open source, meaning their code is available for developers to implement in their own applications, others like Midjourney are proprietary. If you are curious about how you can generate your own images, here are some of the most popular tools you can use right now.
Built by OpenAI, the original DALL-E was released in January 2021 and was succeeded by DALL-E 2 this year. The newer version generates 4x higher resolution images and also maintains a significantly higher accuracy. While the image generator was initially limited to only 200 beta testers, it was made available to all in September. It uses a dataset of text–image pairs to generate realistic images and art from a description in natural language. After you sign up, you can try out 50 prompts a month.
DALL-E and its sequel DALL-E 2 are built on a tweaked version of GPT-3 (Generative Pre-trained Transformer 3). This is a language-based machine learning model, which uses deep learning to produce human-like text. The original GPT-3 was used for text, but the tweaked version that DALL-E uses is used to generate images through a technique called diffusion modelling.
Midjourney is an independent research lab currently working on an AI image generator similar to the ones on this list, with the goal of expanding the imaginative powers of the human species. The AI image generator is proprietory right now. But it can be accessed through a Discord bot. You’ll have to have a Discord account for that, of course, and will then have to join Midjourney’s Discord server. The outputs mostly seem to be the artsy kind here, unlike DALL-E 2 and Stable Diffusion which seemingly focus on photorealistic images based on your input. You can create your first 25 images for free, but you can stretch the limit to 200 with the Basic Membership subscription.
StarryAI has emerged as one of the most user-friendly AI image models on this list, thanks to its easy-to-use interface. It takes a slightly different approach in that it lets you pick the kind of output you’d like based on presets. The Altair preset produces dream-like images that are closer to abstract art, while Orion prioritises realism, and the Argo one is best for artistic and product renders.
StarryAI also lets you upload a custom image to base your art on if you’ve got something specific in your mind. Another unique thing offered by this AI model is that it’s got dedicated apps on Android and iOS, so you can create and share art on the fly. Keep in mind that you can only generate images twice a day and you’ll have to purchase additional credits if you wish to produce more.
Craiyon (formerly DALL-E Mini)
Formerly called DALL-E Mini, Craiyon is the open-source version of the bigger and significantly more powerful AI DALL-E. Craiyon was first launched in April 2022. The AI image model is the brainchild of Boris Dayma, a Houston-based machine-learning engineer.
DALL-E Mini has no relation to OpenAI and was renamed to Craiyon at OpenAI’s insistence. It saw a huge surge in popularity soon afterwards due to its ability to produce cartoony, meme-worthy images. The output isn’t as realistic as some of the others on the list and there’s a high amount of distortion.
Stable Diffusion is another text-to-image model released by a collaboration of Stability AI, CompVis LMU, and Runway. It has support from EleutherAI and LAION. While you will notice some amount of distortion, as with every generator on this list, Stable Diffusion’s outputs are considerably more realistic than some of the others here. The front end of the image generator is DreamStudio, which you can use right now.
Imagen: Not open to public for now
Google’s been working on its own AI image model called Imagen. While its output is similar to DALL-E 2 and Stable Diffusion, it hasn’t been made available to the public. This will probably not change anytime soon. The company has also developed a video version of the AI model, built upon the photo version, which works by upscaling and “predicting” additional frames.