Discrepancy between prompts and outputs

Hi,

I am using ChatGPT 4o.

The prompt: A clean and minimalistic 3D perspective image of a road rendered diagonally. There is an empty space on the surface of the road where a logo can be inscribed later. The road is the only element in the image, with no additional elements typically related to roads. The background is fully transparent, making it suitable for a business presentation. The overall aesthetic is modern and professional.

Generated this image:

I need to have this road curved to the right, that is, it starts on the upper left part of the image, goes down, then curves right (from the image viewer persepctive). The road ends on the right border of the image.

I’ve tried nearly a hundred variations of the following prompt, often clearing the memory:

“Render an image exactly as described. A clean and minimalistic 3D perspective image of a road rendered in a perspective. The road has just one lane. There are the typical white lane markings. The road and the lane markings are the only elements in the image, with no additional elements typically related to roads. The road should start, at the left top, pointing up. It should go down in a straight path until its middle, when it curves right. Then it goes straight to the right. The road ends at the right bottom of of the image, pointing right. The viewer of the image sees it upfront and close to the road. The background is fully transparent, making it suitable for a business presentation. The overall aesthetic is modern and professional.”

It never gets anything even close to that. It generates multiple lanes, it mirrors the image simmetricaly, it adds buildings and other elements, it makes the road curve multiple times, and it never ever makes the end of the road pointing right.

I tried reasoning with it, asking how many lanes it recognizes in the image, I tried asking about its coordinate system, I uploaded an image with a simple curved line and instructed it to make the road follow this path, I asked about the curvature of the road vs. the path I sent it, I asked it to follow strictly what I described… It’s been 2 days and nothing works!

It is really frustrating the lack of consistency and the fact that it can’t check its own image output against the requirements.

I am doing something wrong? Can I improve my prompt or strategy? Is ChatGPT 4o/Dall-E 2 uncapable of delivering consistently a result that sticks to the prompt?