Site icon Wonderful Engineering

This New Tool By OpenAI Can Create Realistic Images From Just Text Input

OpenAI Just Revolutionized Digital Image-Making

The process of creating and modifying photorealistic digital photos is about to get way simpler.

OpenAI, a Microsoft-affiliated artificial intelligence firm based in San Francisco, recently announced the development of an A.I. system that can take a description of an object or scene and automatically generate a very realistic image of it. In addition, the system enables users to change the image using simple tools and text modifications easily.

However, DALL-E 2 is far from flawless. In complex settings, the technology occasionally fails to render details. For example, it’s possible that some lighting and shadow effects are off or that the borders of two items that should be different have merged. It’s also not as good as some other multimodal A.I. applications at comprehending “binding properties.” 

The program is called DALL-E 2, and it is an improved version of a system first introduced by OpenAI in early 2021, known as DALL-E. The initial DALL-E could only render pictures in a cartoonish style, frequently over a simple background. The new DALL-E 2 can produce high-resolution photos with photo-quality backdrops, depth-of-field effects, realistic shadows, shading, and reflections.

Aris Konstantinidis, an OpenAI employee, used the tool to create images of a panda riding a motorcycle wearing a red bandana.

Another OpenAI employee, research and product SVP Mira Murati, demonstrated the technology by using DALL-E2 to build an image from the following text: “macro 35mm film photography of a large family of mice wearing hats cosy by the fireplace.”

DALL-E 2 also simplifies image editing. Draw a box around the portion of the image that needs to be changed and define the modification in natural-language instructions.

Furthermore, OpenAI claims that it has taken steps to limit the software’s capabilities by attempting to eliminate offensive or racist photos from the A.I.’s training data and applying rule-based filters and human content evaluations to the images generated by the A.I.

The tool is rolled out to a limited number of users via a waitlist. We’re excited to get our hands on the tool and put it to the test.