Subject: Image Generation Issue – Model Not Following “Clean-Shaven” Instructions
Hello OpenAI Support Team,
I am experiencing a recurring issue with image generation where the model does not follow explicit instructions to make a character completely clean-shaven (no beard, no mustache, no stubble, no facial hair at all).
Issue Details:
• I have attempted multiple refinements to ensure a male character is clean-shaven, using clear and explicit wording such as:
• “The man has NO beard, NO mustache, NO stubble, and NO facial hair whatsoever. His face should be as smooth as if he has never grown facial hair.”
• Despite this, the model continues to generate images with beards and mustaches even when facial hair is explicitly prohibited in the prompt.
• Even when requesting an edit to remove facial hair, the model often fails to properly erase it or reintroduces it in new generations.
Steps I Took to Fix It:
1. Used multiple variations of wording to reinforce the request.
2. Ensured there were no conflicting details in the prompt.
3. Attempted progressive refinements, but the issue persists.
Expected Behavior:
• The model should accurately interpret and execute the prompt when explicitly instructed to make a character clean-shaven.
• The inpainting/editing function should fully remove facial hair when requested.
Request for Improvement:
• Enhance the model’s ability to recognize and follow instructions regarding facial hair.
• Improve accuracy in inpainting edits when removing specific unwanted elements from an image.
I appreciate your time and look forward to any improvements that can be made to fix this issue. Please let me know if any additional information is needed.
There are still several weaknesses in the image generator. One of them, which many stumble over, is that DALL·E and even GPT itself do not process negations well. Whatever is mentioned in the prompt often ends up in the image. Words like “no” “not,” “without,” or “never” don’t work as expected. what you mention, will mostly end up in the image.
Instead, you have to practice some “DALL·E kung fu.” Rather than blocking something, it’s better to emphasize what you actually want in the image.
It’s also always useful to share the prompt that didn’t work. Others can then experiment with it.
This is the right DALL·E kung fu.
Here is a page that lists many known weaknesses, and some fixing tips, and there’s also a link to a shortened version.
This was not in your question, but I prefer to inform you in advance since you are generating portraits.
Especially in portrait shots, but also in other contexts, DALL·E is still producing errors. These are for example, white spots in the eyes, but often also on the edges or in the hair. This is a flaw in the current model and cannot be easily fixed. The only option is to generate multiple images, and one of them might be generated without the error.
This can only truly be corrected with a model update and is not dependent on your prompt.
