“Fabulous” DALL-E 3 AI Images Allegedly Leaked Online
OpenAI, creators of ChatGPT, are also behind the famous DALL-E image AI. Now sample images of its latest but still-closed closed test version have supposedly been leaked.
If they’re really from the newest version of the visual-rendering AI, these new images are truly something else according to those who’ve seen their original test server and its collection.
One YouTuber, MattVid Pro AI, posted the video below just yesterday with a title that you could interpret as honest, or just typical YouTube clickbait fluff. It’s called “We’ve NEVER seen Image Generation This Good!”
According to this YouTuber, the images he showcases as examples originally came from a user on his Discord channel who claims to be one of only a few hundred (400 to be exact) people worldwide who’s seen “DALL-E 3’s” test server files.
Further dipping into what might be true or just a bit too much hyperbole, MattVideo AI exclaims,
“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,”
Those are some pretty heady statements, but let’s take a peek at some of the alleged leaked visuals to see how they compare:
As you can see, the images above (assuming they’re really from DALL-E 3) are pretty damn good for AI work, and most strikingly, they do a remarkably precise job of rendering text well.
As anyone who’s worked with AI image rendering knows, clean, legible text generation is one of its major weaknesses.
As for claiming that the visuals are better than anything seen so far, or stunningly superior to even Midjourney, meh.
The Grand Theft Auto V mockup, for example, looks good superficially, and its title text looks very authentic, but just a quick scan at the imagery on the cover screams AI. Overall it’s hardly mind-blowing by the current standards of consumer AI.
The same goes even for the Shrek image. A quick glimpse at the people surrounding his mud puddle shows deformed hands. This is another typical AI weakness and Midjourney 5 at least works much better than this for avoiding it.
Overall, the other supposedly leaked images shown here are all great but except for the text rendering quality, nothing in them stands way above what we’ve already seen.
Matt adds that these test visuals come from a possible DALL-E 3 version that still isn’t finalized and lacks any safety features against violent or explicit imagery.
He adds, “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,”
Matt also goes as far as claiming that the AI sometimes adds in extremely violent images even without being prompted to do so.
The YouTuber is claiming that we’ll see a public release Of DALL-E 3 by the end of the year and says that he thinks it could “crush everything else out there.”
I’m not convinced. Even if the visuals really are DALL-E 3 leaks from OpenAI’s servers, (and they very easily could be), they’re far from being revolutionary compared to what we already see daily.
Of course, to avoid sounding a bit too absurdly jaded too early, it’s also worth noting that even current AI image technology really is revolutionary compared to anything that existed up to just a year ago.
The latest current version of DALL-E, DALL-E 2, is about a year old now. It used to be the best of the bunch among major AI image rendering platforms right up until Midjourney v5 blew it out of the water with its frequent hyper-realism.
Even Stable Diffusion seems to be better than DALL-E at creating visuals. Thus, the idea of OpenAI working on a much better version of its platform is definitely believable.