OpenAI has given its artificial intelligence image generator, DALL-E 3, its first major upgrade since its launch last year It can now edit, delete, and adapt any part of an image

Integrated into the ChatGPT AI chatbot, DALL-E 3 can create a wide range of images in landscape, portrait, and square formats to create cartoon and realistic depictions

The new update allows users to select any portion of the generated image and tell ChatGPT what to do with that space In a quick test, I had it remove two random moons from an Earth photo

Such AI image editing, also known as inpainting, is already available in Microsoft's DALL-E implementation of Copilot; Midjourney has had inpainting capabilities since last year, and Adobe Firefly adapts image style reference feature

The OpenAI implementation relies heavily on ChatGPT understanding your request I found during testing that it really struggled when asked to add something, but was good at removing something from a scene

Particularly useful is changing the text generated on the image; DALL-E 3 is much better than its predecessor at adding text, but still makes mistakes regularly, such as adding extra characters in the middle of words or missing a character altogether

Inpainting allows you to select that part of the image, highlight the incorrect letter, and give ChatGPT the correct spelling This is great for posters and flyers

The actual process of image editing is quite simple Open the generated image, click on the inpainting tool, draw the area you want to change, and tell ChatGPT to do it

The problem is that ChatGPT can be a bit lazy and claim to have added an object when no object was added If you respond, "No, it did not," ChatGPT will not retry and will regenerate the image entirely

When it works, it is really impressive and fast In one example, we had it create an image of a cat on the moon It looked great, but I decided to add two moons I highlighted them both and removed the random moon and it did exactly that

However, when I tried to place the lunar lander next to the cat, it was either rejected outright or an extra planet was added behind the lander

Overall, the implementation of inpainting is very good, but there is much to be done to make it more useful For example, OpenAI needs to ensure that ChatGPT actually does what is requested