As reported, DALL-E 2 has been able to generate photorealistic faces from the beginning, but those responsible for the project did not allow them to be circulated outdoors due to the risk of creating deepfakes and false content. So what has changed to allow researchers to share auto-generated faces? OpenAI has implemented several changes to improve the security of this technology.
DALLE-2 can generate faces, but with these limitations
To reduce risks, OpenAI claims that DALL-E 2 now prevents researchers from being able to generate photorealistic faces from any “realistic face.” In addition, the platform rejects any attempt to generate the faces of celebrities and public figures. It should be noted that previously it was not allowed to generate faces of prominent political figures, but after this update, the limitations include any famous personality.
At all times, we speak of researchers because, unlike DALL-E Mini, DALL-E 2 is currently in a closed testing phase, so it is not yet available to the general public. Even those who have access to this technology still have certain limitations. However, numerous examples of how it works can be found on social networks. And some publications let us see what photorealistic faces generated by this technology look like.
Danielle Baskin, via the Twitter account @djbaskin_images, describes how she has used DALL-E 2 to generate photorealistic faces with different makeup designs. “I added the phrase ‘triangular geometric neon yellow’ to my prompts and got these amazing hair and makeup ideas,” he says, adding “when you look into the eyes [at the generated images], you know it’s artificial intelligence.”
An example of how DALL-E 2 can be used to generate images from another image can be seen in a Tweet by Ken Kocienda. Starting from Johannes Vermeer’s famous painting ‘The Girl with the Pearl Earring’, this user asked to generate a studio scene, with photographers, cameras on tripods, and lots of lights, that yes, in “extremely detailed oil painting”. The surprising result highlights the potential of this type of constantly evolving tool.
Behind DALLE-2 is GPT-3, a powerful language model presented in 2020 capable of generating various types of content through written prompts, and a “diffusion” model, which starts with a pattern of random dots and gradually alters that pattern to generate an image when it recognizes specific aspects. Other promising proposals in this field include Google’s recently unveiled Parti, which uses an “autoregressive model” and allows for longer text inputs.