How Can I Opt-Out of DALL·E 3’s Automatic Prompt Enhancements in the API?

Below is a summary of the behavior you're describing regarding DALL·E 3's prompt modification and the “opt-out” trick:

---

Automatic Prompt Enhancement

• Default Behavior:
  DALL·E 3 automatically rewrites and elaborates on your input prompt for safety and to enhance image quality. Even if you provide a very simple prompt (e.g., “images of ducks”), the system augments it with more detail. The revised version is returned in the response under a field (commonly seen as revised_prompt).

• Rationale:
  This design is intended to improve the generated images by introducing additional descriptive elements, which usually leads to higher quality outputs.

---

Opt-Out via Prompt Magic

• How It Works:
  It isn’t currently possible to disable this feature through any API setting. However, there is a workaround: by prepending a specific phrase to your prompt you can instruct the system not to add any extra detail. For example:

  PLAINTEXTCopied
  I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:

• Use Case:
  This is especially useful for developers or users who want to inspect the “raw” output or compare how the augmented prompt affects the generated image.

---

Some Developer Insights

• Single vs. Multiple Outputs:
  There’s also mention in the community that with DALL·E 3, you can only request one image per API call (with the option to send parallel requests for multiple images). In contrast, DALL·E 2 allowed up to 10 images with the n parameter.

• Response Inspection:
  Checking the response object (which includes the revised prompt) is a helpful way to ensure that the system’s modifications are in line with your expectations. If the changes degrade the prompt’s effectiveness for your particular application, you can use the opt-out trick mentioned above.

---

Final Thoughts

What you’re observing aligns with documented behavior: DALL·E 3 is designed to enhance prompts by default to optimize safety and quality. The “inception”-like approach you speculated about (i.e., reprocessing the prompt via an internal model) is a reasonable hypothesis, but the main takeaway is to always review the revised_prompt field. If the automatic enhancements aren’t beneficial to your application, prepend the designated phrase to your prompt to prevent the additional modifications.

This approach has been discussed widely within the community, and several threads (like those referenced from Paul Bellow and others) offer tips on working around or leveraging this feature.

If you have further questions about the API or need more detailed guidance on structuring prompts, feel free to ask!