Rumored ‘DALL-E 3’ Images Leaked and it Can Finally Generate Text Right

Images allegedly generated with a brand new version of OpenAI’s DALL-E, possibly DALL-E 3, have been leaked online and they are exceptional.

YouTuber MattVidPro AI posted the above video yesterday titled “We’ve NEVER seen Image Generation This Good!” and in it claims the images came from a user on his Discord channel who is one of just 400 people in the world to have access to an exclusive testing server for the latest version of DALL-E.

“We’re truly glimpsing into the future of AI image generation…I promise you, you have not seen any AI image generations that are this good. Midjourney cannot compete at this level,” he says.

Shrek being discovered in a bog.
One of the leaked images showing Shrek being discovered in a bog.

DALL-E 2 is over a year old and was the best in class until Midjourney and even Stable Diffusion seemingly overtook it. But Matt claims that OpenAI has not stopped working on DALL-E and may be about to unleash an image synthesis model that is a cut above anything else on the market.

A Step Up in AI Image Generation

One of the striking features of the images that Matt shows in his video is the model’s apparent ability to nail text in an image, something that AI image generators are notoriously bad at.

An invented novelty box with near-perfect text and spelling.
Grand Theft Auto V
Grand Theft Auto V.

One of the more incredible examples is a product shot of the video game Grand Theft Auto V which at first glance is a near-perfect representation of one of the world’s best-selling games in a PS4 disc box.

“Most current AI image generation models can’t come close to spelling text this coherently,” adds Matt.

Animals made from cheese and ham while ah hungry dog looks on.
Animals made from cheese and ham while a hungry dog looks on.

According to Matt, this latest DALL-E model has not been finalized and all safety features have been removed; meaning violent and other upsetting imagery can currently be generated.

“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,” says Matt.

The YouTuber adds that extreme violence can apparently just show up in images even without asking for it in the prompt.

“They’ve clearly got some fine-tuning and some safety work to do,” says Matt.

Matt goes on to predict that “DALL-E 3” will be released by the end of the year and it might “crush everything else out there.”

PetaPixel has approached OpenAI for comment.