It seems there has been an improvement in the accuracy of dall-e3 via Bing, but since this can't be in the image generator itself(which hasn't been updated to my knowing), they must have improved the layer in between, the way chatGPT is used to generate middle-prompts based on what prompt the user inputs.

Pic related was impossible to create just a month ago. It would just downright refuse to show any result (the prompt being innocent, the image was just not shown).