Issues when using dall-e-3 model


I am using dall-e-3 via API, and find some issues:

  1. Sometimes the image will filled by many objects, even I have add the text “Do NOT put too many objects to the image.”. Why?

  2. There are always typos in the image. For example, “Access” is misspelled as “Acess”, and so on.

  3. I am using 1792x1024 dimension. It seems that sometimes the image is cropped and some objects are cut in half.

How to solve these issues?

Thank you

1 Like
  1. Negative prompts are hard for an LLM to achieve - ie telling it NOT to do something. When it adds a lot of items/people, it usually means your prompt is not detailed enough, and it’s filling in all the space it can. Be specific about number of items/people in the scene, so it knows what you want. It’s not a mind reader (yet!)

  2. That it can do text at all is amazing, but yes, it still has a lot of flaws and likely shouldn’t be used in production settings.

  3. This is sometimes the prompt (a single word which can be hard to pin down)… The DALLE team knows about this (and other) problems and are diligently working on them, I’ve heard!