OpenAI has expanded access to its DALL-E 3 AI image generator, and the company is offering Plus and Enterprise customers access to new DALL-E 3 models through direct integration with ChatGPT. The company said it has implemented various security measures to reduce the generation of malicious images for widespread distribution.
OpenAI first released DALL-E 3 last month and the company showed how the model improves on its predecessor DALL-E 2 by allowing users to type long text prompts while leveraging ChatGPT to get the image generator result.
The company said: “The new model produces visually clear and detailed images compared to previous models, as the DALL-E 3 model responds to fully detailed images in addition to presenting complex details such as text, hands and faces.” Horizontal and vertical aspect ratios.
The company added DALL-E 3 to Bing Chat and the Bing Image Generator, making Microsoft's platform the first to provide broad public access to the model, before ChatGPT.
The new template also allows users to create printed text embedded in the generated images. This is especially useful for those who want to create business content (such as marketing materials) or get inspiration for product design.
OpenAI says its new model outperforms several competitors in creating images using AI technology, including Adobe's Firefly 2 and Stable AI's Stable Diffusion.
The company's statement on its guardrails comes in the wake of reports that guardrails posted to reduce harmful images have not been successful.
Users created images of the World Trade Center, and other simple solutions produced similar results, even after Microsoft blocked some prompts.
“We use a multi-layered security system to limit models' ability to create potentially harmful images, including violent or hateful content,” the company wrote. Security checks are done through user prompts and generated images before users see them. "
“We are also working with early adopters and experts to identify and address vulnerabilities in our security systems resulting from the capabilities of the new model,” she added. “Feedback helps us identify edge cases for creating graphic content and evaluate the model. Stress test the ability to create convincingly misleading images.
Other protections are intended to limit DALL-E 3's ability to create content in the style of images of live performers and public figures, and the company has worked to improve demographic representation through images of models.
OpenAI has created a website to promote transparency that showcases the research that led to DALL-E 3, as well as an image classifier that can determine whether an image was generated by its tool with about 95% accuracy. 99%.