OpenAI releases the DALL-E third version
OpenAI has released DALL-E, a generative AI visual art platform, in version three. This version adds extra safety features and allows users to design prompts using ChatGPT.
Text prompts are converted to pictures using DALL-E. However, even DALL-E 2 made mistakes and frequently disregarded certain language. According to OpenAI researchers, the most recent version has far improved context understanding.
DALL-E 3’s connection with ChatGPT is a new feature. With ChatGPT, a user doesn’t need to think of a thorough prompt on their own; instead, they can ask the chatbot to generate one, and it will compose a paragraph (longer sentences work best for DALL-E) for DALLE 3 to follow. If other people have special ideas for DALL-E, they can still utilize their prompts.
Aditya Ramesh, principal researcher and head of the DALL-E team, asked ChatGPT to assist him in creating a logo for a ramen restaurant in the mountains during a demo for The Verge. After that, ChatGPT penned a longer prompt, and DALL-E suggested four possibilities. My favourite image was of a mountain with pickled eggs on the ground like garden stones, broth cascading down like a waterfall, and ramen snowcaps; nonetheless, it appeared more like an illustration for some lovely merchandise than a typical restaurant logo. According to OpenAI, this interaction with the chatbot makes it possible for more individuals to produce AI art because it eliminates the need for highly developed prompt writing skills.
DALL-E was first made available in January 2021, ahead of rival text-to-image generative AI art systems from Midjourney and Stability AI. By the time DALL-E 2 was made available in 2022, OpenAI had established a queue to manage access to the platform following complaints that DALL-E was biased and capable of producing lifelike sexual imagery. In September of last year, the business launched DALL-E 2 to the public and eliminated the waiting.
Research laboratories and its API service will receive this new version of DALL-E in the fall after it is initially made available to ChatGPT Plus and ChatGPT Enterprise subscribers in October. While OpenAI intends to deploy DALL-E 3 gradually, the company has not stated when a free public version will be made available.
According to OpenAI, a lot of effort went into developing strong safety features for DALL-E 3 in order to avoid the production of offensive or potentially malicious images. According to OpenAI, it relied on input classifiers, a technique for teaching language models to disregard specific words in order to avoid explicit or violent prompts. It collaborated with external red teamers, a group that purposefully attempts to breach a system to evaluate its safety. Additionally, DALL-E 3 won’t be able to replicate pictures of famous people unless the prompt expressly states their names.
Policy researcher Sandhini Agarwal of the company expressed her “high confidence” in the company’s safety protocols, although she noted that the model is not flawless and is always improving. In an email, OpenAI staff stated that DALL-E 3 has been trained to refuse to produce images that mimic the works of living artists. As opposed to DALL-E 2, which can imitate select artists’ works of art when asked.
In addition, OpenAI will let artists remove their work from further iterations of text-to-image AI models, presumably to prevent legal action. Through a form on its website, creators can submit an image to which they are the rightful owner and request that it be removed. In the future, DALL-E will have the ability to filter out results that resemble the picture and style of the artist. Artists filed lawsuits against DeviantArt, Stability AI, and Midjourney, rival DALL-E companies, alleging that they were utilizing their copyrighted works as training data for their text-to-image models.
