OpenAI’s DALL-E 2 may now generate faces
OpenAI’s DALL-E 2 is now allowed to generate faces after all.
Since the launch of DALL-E 2 in mid-April, more and more testers have had access to OpenAI’s image AI. As of this week, they are also allowed to generate faces, a feature the company had previously blocked for security reasons. OpenAI feared fraud attempts with fake images of celebrities and politicians.
Faces of non-existent people are fine
DALL-E’s testers are now allowed to generate and publicly share images of faces, thanks to two new security measures designed to minimize the risk of misleading content.
According to OpenAI, its system can now successfully block attempts to generate images of people who are in the public eye. The generated images only show faces of non-existing people. Generating human faces opens up new application scenarios for DALL-E 2, especially in an artistic context, as the following examples show.
OpenAI has decided today that it’s safe for us to start sharing images containing realistic faces. DALL-E2 is actually excellent at faces.
From here on, this feed will largely be machine dreams of people that never existed.
– Patrick Clair (@PatrickClair) June 21, 2022
It remains to be seen how OpenAI will handle photorealistic human faces in a documentary context. Here, prominence would not be a prerequisite for social impact, as it is with photos of fictional children from a war zone, for example. According to the current content guidelines, such an image composition would probably fall under “deception” and be prohibited.
- MIXED.de ohne Werbebanner
- Zugriff auf mehr als 9.000 Artikel
- Kündigung jederzeit online möglich
DALL-E 2 safety: First interim result is encouraging
At the end of May, OpenAI published a first interim report on the compliance of DALL-E 2 after more than three million images had been generated.
The result was very positive: only 0.05 percent of the generated images were identified as potential violations of OpenAI’s content guidelines. Of this 0.05 percent, only 30 percent were actually penalized – that’s about 450 images out of three million.
DALL-E 2 is also likely to be a commercial success. Already now, the interest in DALL-E 2 is significantly greater than in the text AI GPT-3 or the programming AI Codex / Copilot.