The unreleased generative imaging tool doesn’t shy away from nudity and gore, and caught the attention of MidJourney.
OpenAI has been privately testing a new version of its generative artificial intelligence (AI) imaging model for the past few months, and early samples leaked by YouTube user MattVidPro show that the model outperforms previous image generators.
"It's very exciting, it blows the mind of anything we've seen before, it's crazy," Matt said in a preview posted on YouTube. "Midjourney can't compete at this level, I didn't even think Midjourney 6 could compete at this level."
Don't expect to try it anytime soon, though. Access is extremely limited.
The unreleased model is likely an upgraded version of DALL-E 2 and is being tested via an invite-only preview within ChatGPT-4. Matt said only about 400 people worldwide have access to this new OpenAI image generator.
Although the image samples are limited, they demonstrate the advanced skills of the artificial intelligence. It generates clear images by mimicking the lighting and reflections of real photos, and the model reproduces the details of paintings down to the point where brushstrokes are visible. It also perfectly recreates brand names such as "Snickers" and logos of well-known brands such as Subway on the generated products, and achieves fairly good spelling in the rendered text.
While current image generators struggle to achieve coherent hands, these examples show realistic and properly proportioned hands with more convincing backgrounds than competing AI systems.
OpenAI apparently removed its safety filters to test the full potential of the model. Users said it could generate violent content and nudity without hesitation. However, given OpenAI's stance on NSFW content, it is highly unlikely that an official public version will be released to such a standard.
Matt observed, “This is OpenAI, so you can bet they’ll shut it down like they do with everything before release.”
Some experts criticized OpenAI for “simplifying” its model to avoid potential controversy. Some studies even showed that ChatGPT trained by OpenAI had a strong political bias in its output.
Still, the consistent quality shown in the samples is a leap forward that underscores OpenAI’s ongoing efforts to improve the capabilities of generative AI. The company may reveal more about its progress later this year, especially if the areas of image recognition and generation help improve the robustness of its star product: the multimodal GPT-4, which can understand text, images, and sketches in a single prompt.
Currently, the technology is still limited to closed testing with a very small number of users.
As models continue to improve, the line between artificial and real becomes more blurred. While this excites many, concerns about misuse will persist, and building this technology responsibly remains a pressing challenge.