Mon. Jul 15th, 2024

DALL-E 2, OpenAI’s AI, can create photorealistic faces of fictional persons.

By b0oua Jun 8, 2024

At the beginning of this year, DALL-E 2 demonstrated to us its remarkable capability to generate images from any text, which is a significant advancement in comparison to the prior generation. However, throughout all of these months, this artificial intelligence developed by OpenAI has been subject to a significant limitation. Previously, the conditions of use barred researchers who had access to it from sharing photos of faces that were generated automatically. However, this limitation has recently been removed.

The fact that DALL-E 2 has been able to manufacture photorealistic faces from the very beginning is something that our colleagues at Genbeta have reported. However, the individuals responsible for the project did not permit its circulation outside of the company because of the possibility of producing deep fakes and misleading content. So, what has changed in order for researchers to be able to share faces that were generated automatically? An assortment of modifications have been done by OpenAI in order to enhance the safety of this technology.

OpenAI has stated that DALL-E 2 has been updated to restrict researchers from producing photorealistic faces from any “realistic face.” This change was made in order to lessen the hazards. A further point to consider is that the site does not tolerate any attempts to manufacture the faces of prominent figures and celebrities. Previously, it was not permitted to generate faces of major political leaders; however, with this update, the constraints have been expanded to cover any notable personality. It is important to highlight that this change was made.

At all times, we discuss researchers since, in contrast to DALL-E Mini, DALL-E 2 is still in a closed testing phase, which means that it is not yet available to the general public. Even members of the population that have access to this technology are nevertheless subject to certain constraints. On social networks, however, there are a great number of examples that demonstrate how it operates. In addition, there are publications that provide us with the opportunity to view the photorealistic faces that are produced by this technology.

Danielle Baskin, through the Twitter account @djbaskin_images, describes how she has utilized DALL-E 2 to construct photorealistic faces with varied cosmetics designs . “I added the phrase ‘neon yellow geometric triangle’ to my prompts and received these fantastic hair and cosmetics suggestions,” she adds, adding “when you look into the eyes [of the created images], you know it’s artificial intelligence.” “.

Ken Kocienda’s tweet provides an illustration of how DALL-E 2 can be utilized to generate images from another image. This example can be seen in the tweet. Beginning with the well-known artwork “Girl with a Pearl Earring” by Johannes Vermeer, this user requested that a studio setting be generated, complete with photographers, cameras mounted on tripods, and a large number of lights, according to the user’s description of the “extremely detailed oil painting.” The outcome, which is unexpected, demonstrates the potential of instruments of this kind that are constantly undergoing their evolution.

GPT-3, a powerful language model that was introduced in 2020 and is capable of generating various types of content through written prompts, is the driving force behind DALLE-2. Additionally, there is a “diffusion” model that begins with a pattern of random dots and gradually alters that pattern to generate an image when it recognizes specific aspects. There are also other intriguing suggestions in this area, such as the Parti algorithm that was just recently presented by Google. This algorithm makes use of an autoregressive model and permits longer text entries.

In the beginning, DALL·E 2 was a research project, and it is currently available in beta form. Some of the safety precautions that we have developed and are continually working to improve include:

Preventing generations from being detrimental
We have restricted the capability of DALL·E 2 to produce images that are aggressive, hateful, or even adult in nature. Through the elimination of the most explicit content from the training data, we were able to reduce the amount of exposure that DALL·E 2 had to these notions. A number of cutting-edge procedures were also utilized in order to prevent the development of photorealistic versions of the faces of real people, particularly those of public figures.
Study up on ways to make things safer.

Reducing the amount of misuse
In accordance with our content policy, users are not permitted to generate anything that falls into any of the following categories: violent, adult, or political. In the event that our filters identify text prompts and image uploads that may be in violation of our regulations, we will not automatically produce images. As an additional measure to prevent misuse, we have both automated and human monitoring mechanisms in place.
Please read the content policy, which will open in a new window.

Learning-based phased deployment of the system
When it comes to developing and implementing artificial intelligence in a responsible manner, learning from real-world applications is an essential component. We started off by providing a preview of DALL·E 2 to a select group of people that we relied on. Based on our understanding of the capabilities and limitations of the technology, as well as our confidence in our safety mechanisms, we gradually increased the number of users and eventually made DALL·E available in beta form in the month of July in the year 2022.

We have high hopes that DALL·E 2 would enable individuals to express themselves creatively and with confidence. DALL·E 2 provides us with the ability to comprehend how sophisticated artificial intelligence systems perceive and comprehend our reality, which is of utmost importance in our endeavor to develop AI that is beneficial to humanity.

The following individuals were included in the list: Jeff Belgum, Dave Cummings, Jonathan Gordon, Chris Hallacy, Shawn Jain, Joanne Jang, Fraser Kelton, Vishal Kuo, Joel Lehman, Rachel Lim, Bianca Martin, Evan Morikawa, Rajeev Nayak, Glenn Powell, Krijn Rijshouwer, David Schnurr, Maddie Simens, Kenneth Stanley, Felipe Such, Chelsea Voss, and Justin Jay Wang.

The following individuals were included on the list: Steven Adler, Lama Ahmad, Miles Brundage, Kevin Button, Che Chang, Fotis Chantzis, Derek Chen, Frances Choi, Steve Dowling, Elie Georges, Shino Jomoto, Aris Konstantinidis, Gretchen Krueger, Andrew Mayne, Pamela Mishkin, Bob Rotsted, Natalie Summers, Dave Willner, and Hannah Wong.

A special thank you goes out to everyone who contributed to and offered input on this release: Sandhini Agarwal, Sam Altman, Chester Cho, Peter Hoeschele, Jacob Jackson, Jong Wook Kim, Matt Knight, Jason Kwon, Anna Makanju, Katie Mayer, Bob McGrew, Luke Miller, Mira Murati, Adam Nace, Hyeonwoo Noh, Cullen O’Keefe, Long Ouyang, Michael Petrov, Henrique Ponde de Oliveira Pinto, Alec Radford, Girish Sastry, Pranav Shyam, Aravind Srinivas, Ilya Sutskever, Preston Tuggle, Arun Vijayvergiya, Peter Welinder–6663fe27d69e0#goto7792–666437d99ad9f#goto7798—-saudi-arabia-632293553

By b0oua

Related Post

Leave a Reply

Your email address will not be published. Required fields are marked *