OpenAI is reportedly developing DALL-E 3, an upgraded version of its text-to-image generator.
According to YouTuber MattVidPro AI, around 400 people currently have access to a private test server for DALL-E 3, and there have been several leaked images of its work.
A video shared by MattVidPro AI showed how DALL-E test versions have progressed from May to July.
This evolved version of DALL-E seems to be able to generate photo-real images and use specific techniques, such as brush strokes or pencil drawing.
The unreleased version of DALL-E can also generate images depicting gore, blood, and copyrighted content, according to MattVidPro AI – he provided some examples of photos involving a Snickers bar and Subway.
It’s also excellent at producing typography, an Achilles heel of past and present text-to-image models.
The YouTuber described the tool as “extremely exciting. This blows anything we’ve seen before out of the water. It’s insane.”
Healso stated that “Midjourney cannot compete at this level. I don’t even think Midjourney version six would be able to compete at this level.”
Commenters on the video highlight MidJourney’s poor ability to adhere to prompts, which DALL-E 3 supposedly addresses.
He continues, “You can do exact copies of copyrighted artworks and materials. It will do copyrighted characters perfectly, logos perfectly, completely accurate depictions of celebrities, and full-out depictions of extreme violence” and goes on to say the model sometimes produces inappropriate imagery without being explicitly programmed to do so.
“They’ve clearly got some fine-tuning and some safety work to do,” says Matt.
Comparisons to other text-to-image AIs
Comparisons between the new DALL-E and Midjourney’s latest model showed OpenAI’s content to be superior, with better adherence to given prompts and improved image quality.
However, the DALL-E 3 upgrade is still under development, and its official release date remains uncertain. Of course, this YouTube video is designed to hype up DALL-E 3, and the public version will likely be very different from any ‘secret’ test versions.
Once released, DALL-E 3 will likely feature several restrictions, such as neutering images involving blood, gore, copyrighted content, and reproductions of works of art.
Will DALL-E 3 put OpenAI back on top of the image generation AI market?