DALL-E 2 turns written prompts into editable images
OpenAI previews DALL-E 2, a system that generates high-resolution images from text and can extend or alter pictures while preserving visual relationships.
OpenAI introduced DALL-E 2 today, showing a research system that creates images from written descriptions. It can combine subjects, attributes and artistic styles, produce variations of an existing picture and edit selected regions using natural-language instructions.
The new model produces higher-resolution results than the original DALL-E and uses a process related to OpenAI's CLIP image-text system. Access is limited while the company studies risks and gathers feedback from a controlled group of users.
Some prompts yield strikingly coherent lighting, texture and composition, making the tool feel closer to a visual collaborator than a search engine. It can also misunderstand relationships and reproduce biases from training data. OpenAI is restricting violent, sexual and political content and has removed certain images from training. Those safeguards will be tested as access expands, especially when generated pictures can look persuasive without documenting any event that actually occurred.
Source: OpenAI ↗