The company unveiled the original DALL-E in January 2021, with the tool impressing both AI experts and the public with its ability to turn any text description (or prompt) into a unique image. Since then, a number of other text-to-image systems have been created that rival the speed and quality of DALL-E. Other systems, like Midjourney and Stable Diffusion, are much easier for anyone to access, drawing attention away from OpenAI’s own offering.
Open access AI art image generators have drawn attention away from DALL-E
OpenAI, which has received substantial funding from tech giant Microsoft, has been cautious about the public release of DALL-E. Experts note that the ability of text-to-image systems to produce nonconsensual nudes and photorealistic images is potentially damaging — fodder for harassment, propaganda, misinformation, and more. There are also issues of bias. Because text-to-images systems are trained on vast datasets of images scraped from the internet, they replicate unequal aspects of society. Ask a system to draw a picture of a CEO, for example, and it will typically produce an image of a white man.
OpenAI has taken a number of measures to combat these effects, including filtering out sexual and violent imagery from its training data and refusing to generate images based on similarly explicit prompts. However, the company has also been criticized for what some see as an overly restrictive or clumsy approach to mitigating harm.
Emad Mostaque, who helped develop rival text-to-image AI Stable Diffusion, said, for example, that it was an “asshole move” for OpenAI not to generate images from words like “Ukraine” and “Odesa.” (Presumably these terms are censored because of their potential to create misinformation during an ongoing war). Others have criticized the company’s attempts to fix bias as “hacky.” For example, DALL-E invisibly inserts phrases like “Black man” and “Asian woman” into user prompts that do not specify gender or ethnicity in order to nudge the system away from generating images of white people. (OpenAI confirmed to The Verge that it uses this method.) This does mitigate bias in DALL-E’s output, but some users have noted it also creates unwanted imagery that doesn’t match their instructions.
OpenAI is confident it’s mitigated DALL-E’s potentially harmful effects
In its blog post today, OpenAI said it was satisfied with the improvements it’s made to its safety systems and that this will help offset potential harms as DALL-E becomes more accessible. “In the past months, we have made our filters more robust at rejecting attempts to generate sexual, violent and other content that violates our content policy, and building new detection and response techniques to stop misuse,” said the company.
The firm also said it’s testing an API for DALL-E that would allow companies to build their own apps and plug-ins using the system’s output. This would make it much easier for OpenAI to commercialize DALL-E’s output, potentially combining the system with tools used by illustrators and designers, for example.
Anyone who signs up for access to DALL-E will get 50 credits free and then 15 more free credits every month after that. Each credit can be used to generate a single image, a variation of an image, or for “inpainting” and “outpainting” (editing the contents of an image or extending an image beyond its existing boundaries). Additional credits can be bought in blocks of 115 for $15. OpenAI says some 1.5 million DALL-E users are generating over 2 million images every day.