DALL-E Mini’s viral moment doesn’t just herald a new way to make memes. It also provides an early look at what can happen when AI tools that make imagery to order become widely available, and a reminder of the uncertainties about their possible impact. Algorithms that generate custom photography and artwork might transform art and help businesses with marketing, but they could also have the power to manipulate and mislead. A warning on the DALL-E Mini web page warns that it may “reinforce or exacerbate societal biases” or “generate images that contain stereotypes against minority groups.”
DALL-E Mini was inspired by a more powerful AI image-making tool called DALL-E (a portmanteau of Salvador Dali and WALL-E), revealed by AI research company OpenAI in January 2021. DALL-E is more powerful but is not openly available, due to concerns that it will be misused.
It has become common for breakthroughs in AI research to be quickly replicated elsewhere, often within months, and DALL-E was no exception. Boris Dayma, a machine learning consultant based in Houston, Texas, says he was fascinated by the original DALL-E research paper. Although OpenAI did not release any code, he was able to knock together the first version of DALL-E Mini at a hackathon organized by Hugging Face and Google in July 2021. The first version produced low-quality images that were often difficult to recognize, but Dayma has continued to improve on it since. Last week he rebranded his project as Craiyon, after OpenAI requested he change the name to avoid confusion with the original DALL-E project. The new site displays ads, and Dayma is also planning a premium version of his image generator.
DALL-E Mini images have a distinctively alien look. Objects are often distorted and smudged, and people appear with faces or body parts missing or mangled. But it’s usually possible to recognize what it is attempting to depict, and comparing the AI’s sometimes unhinged output with the original prompt is often fun.
The AI model behind DALL-E Mini makes images by drawing on statistical patterns it gleaned from analyzing about 30 million labeled images to extract connections between words and pixels. Dayma compiled that training data from several public image collections gathered from the web, including one released by OpenAI. The system can make mistakes partly because it lacks a real understanding of how objects should behave in the physical world. Small snippets of text are often ambiguous, and AI models do not grasp their meaning in the way that people do. Still, Dayma has been amazed by what people have coaxed out of his creation in the past few weeks. “My most creative prompt was the ‘Eiffel Tower on the moon’,” he says. “Now people do crazy things—and it works.”
