Dear OpenAI Team,
I’d like to provide feedback on my experience with the DALL-E image generation process through ChatGPT, specifically regarding repeated issues with rendering accuracy that led to unnecessary delays.
The main issue I encountered was that the AI consistently failed to remove certain elements from the images despite clear instructions. For example, when I requested an image without wing-like or angelic figures, the AI continued to include these elements, even after multiple corrections. Another example was when I asked for an image of a black hole with no stars near it, but the AI repeatedly generated images containing stars.
These mistakes forced me to request image generation several times, which not only slowed down my progress but also monopolized the system’s resources—resources that could have been available for other users if the image had been accurate from the start.
To improve the experience, I suggest:
- Enhancing the model’s ability to accurately follow specific user instructions (e.g., explicitly excluding certain visual elements).
- Improving the system’s responsiveness to corrections, so it doesn’t generate repeated errors that frustrate users and overload the system.
- Ensuring that initial renders more closely match detailed user prompts to avoid wasting time and resources on multiple iterations.
I believe these improvements would help make the tool more efficient, freeing up capacity for other users and providing a smoother experience overall.
Thank you for your consideration.
Best regards,
Stuff like this work, describe it not using descriptors . Black hole esq spiral on black background
Holy beings floating in air
From my experience naming it to get rid of it don work its best to describe things not using the words you don’t want.
Yes I know your pain.
The image editor cannot reliably converge on a finished image.
I had an image that was approximately 90% correct. I then spent around half a day trying to correct relatively minor issues. When asked to edit one specific element, the system repeatedly altered unrelated elements that were already correct and had not been selected for editing.
This created a never-ending correction loop: fixing one error introduced another; fixing that error changed something previously correct; correcting that change then reintroduced an earlier problem.
Examples included changing established character faces when editing an unrelated area, failing to use a supplied character-face reference, and reintroducing an anatomically impossible third foot after it had already been corrected.
These were not complicated requests. They included things as simple as making two characters look slightly younger, using the exact face reference I supplied for another character, and removing an extra foot.
The result was not simply wasted image generations or inconvenience. A nearly completed image ultimately became unusable and had to be discarded because the editing system could not preserve the parts of the image that were already correct.
This is a fundamental problem for an image-editing product. An editor needs a reliable mechanism to preserve everything outside the requested edit. If asking to change one small area can effectively regenerate or reinterpret unrelated parts of the image, complex images can enter a correction loop from which there is no practical way to reach a finished result.
The initial image generation was good. The editing process made completing that image impossible.
I spent approximately half a day attempting to correct errors, many of which had themselves been introduced by previous edits going outside their requested scope, and ended the process with no usable final image.
OpenAI needs to understand how serious this usability problem is. This is not simply a case of an image being imperfect or a prompt needing refinement. The editing behaviour actively prevented a nearly finished image from being completed.
