OpenAI's DALL-E 3 adds ChatGPT integration, provides a more nuanced approach to text-based...

Alfonso Maruccia

Posts: 755   +259
The big picture: DALL-E is one of the leading AI services designed to generate images from textual prompts. Developed by OpenAI, this machine learning model is continually evolving to offer users more advanced and user-friendly tools for transforming their ideas into uncanny visual content.

OpenAI has announced DALL-E 3, the new generation of its well-known text-to-image generation algorithm. DALL-E 3 can work with nuanced requests and generate "extremely detailed and accurate images," according to the San Francisco-based corporation. It has been built natively on ChatGPT's ML chatbot model.

DALL-E 3 allows users to use ChatGPT as a kind of "brainstorming partner" and refiner of their textual prompts, as explained by OpenAI. Users can ask the chatbot to create images from a simple, one-sentence idea or a complex, detailed paragraph. When given an idea, ChatGPT will automatically generate the most appropriate and "tailored" prompt to feed to DALL-E's text-to-image AI model.

If the resulting image is not quite right, OpenAI states that users can ask ChatGPT to tweak the existing prompt with just a few words. Like previous versions, DALL-E 3 limits the ML model's ability to generate "violent, adult, or hateful" content, although some resourceful users have found ways to bypass these alleged limits in the past.

As an additional measure to prevent "harmful generations," DALL-E 3 has mitigations in place to decline requests asking for images of known public figures. Safety performance has been "improved" through stress-testing sessions conducted by experts, according to OpenAI. Furthermore, the company is researching the best way to help people identify when an image was created with AI.

OpenAI is experimenting with a "provenance classifier," which is a new internal tool for AI image identification. However, OpenAI has not yet shared this tool with its users. DALL-E 3 is also designed to decline requests that ask for an image mimicking the style of a "living artist," OpenAI says. Creators can now also opt out their images from future algorithm training sessions.

OpenAI claims that DALL-E 3 is a significant improvement over DALL-E 2. Even when tasked with the same textual prompt, images generated by the newly-trained algorithm are much more faithful to the user's request.

DALL-E 3 will be available to ChatGPT Plus and Enterprise customers in October, with plans to roll it out to the API and in Labs later this fall. Microsoft, Shutterstock, and other OpenAI partners will likely be among the first to benefit from this improved image-generation technology.

Permalink to story.