For security reasons, OpenAI previously blocked the uploading and editing of images with faces. Now the company is enabling the feature because it believes it can ensure security.
DALL-E-2 users can now upload images of people and edit them with text prompts. The feature was previously available in a beta version, but was initially removed for security reasons. Image upload is possible directly below the DALL-E input mask.
DALL-E 2 lets you upload and edit faces
The fact that the edit function was missing until now and DALL-E 2 was not even allowed to generate faces until a few days ago had to do with security concerns on the part of OpenAI. The company feared misuse, e.g. that the faces of celebrities would be disfigured with DALL-E 2.
OpenAI now appears to have largely addressed these concerns thanks to an improved security system: “With improvements in our safety system, DALL-E is now ready to support these delightful and important use cases – while minimizing the potential of harm from deepfakes,” the company writes.
According to OpenAI, one reconstructive surgeon used the facial editing feature to help his patients visualize possible surgical outcomes in advance. Many would upload images of themselves to test outfits or hairstyles. Filmmakers edit scenes with people in them “to speed up creative processes,” OpenAI said.
OpenAI improves filtering, detection, and response system
An improved filtering system is designed to prevent users from generating sexual, political, and violent content that violates OpenAI’s content policy. In addition, OpenAI has implemented new detection and response techniques to prevent abuse. At the same time, the number of false positives has been reduced by these systems.
As before, users may only upload images to which they own the rights. Images with persons may only be uploaded with the consent of these individuals.
An initial interim security assessment of OpenAI in May 2022 showed that there had been few attempts at misuse. The company is also continuously working to improve the security of DALL-E, e.g., by filtering training data, but this in turn may have other, unexpected side effects.