Question about Dall-E and carrying over prompts

So this is my second post in two days! Hope I’m not getting annoying! Got a question about Dall-E. Me and my brother are toying with ideas for a series of children’s books. The problem I’m running into is this: we got it to create a likeness of a boy that we want to continue with, but we can’t get Dall -E to replicate the boy in future scenes. It just spits out different boys every time no matter what the prompt I put in is. I’ve even inserted pics of the boy we want it to replicate and ask it to replicate it and it doesn’t. Is this just how Dall-E works right now? Or is there a way to get it to replicate images for purposes like the one I described? Thanks for any assistance!

Seeds aren’t really working great yet.

The best you can do is keep the prompt as similar as possible for the boy then at the end of the prompt, change the scene. It’s not 100% reliable, though.

I believe this is something the DALLE team is working on, though, so stay tuned.


Awesome! That’s good to hear. It would be great to have those sorts of capabilities!

1 Like
  1. DALL-E 3 has no image input. It cannot base things on other imagery.
  2. ChatGPT Plus is the only platform that gives a chance of generating images that look the same.
  3. The AI won’t replicate what you input (without great effort), it is told to rewrite prompts.

but the big one:

  1. Only within ChatGPT Plus, the AI has a method to refer to a previous image when it creates a new one. Since this is independent of the prompt, this likely reuses a seed or other embeddings variation that is employed for diversity.


Have one ChatGPT session where you do all the creation. Get the independent “subject” to an appearance that you want by prompting, perhaps 40 words of that, and the style also well-described, perhaps grounded in a named style (although you cannot use 20th Century artist names) that also doesn’t change.


Have the AI reproduce exactly the prompt it sent, using a markdown code block to emphasize that it is code-like, along with printing a gen_id that was returned from the image creation. This emphasizes the text that was employed to the AI. Then have the AI use the exact prompt and gen_id for all future generations in that chat session, specifying in prompt and also referring to the previous image, and only change a “situation” or “plot” section.


Dang! I got a lot of learning to do. Thanks for giving me some direction!

I’m going to use a combination of different approaches, having either templates or gen id’s and GPTs, it’s much more flexible than if each of them were separate.

In addition, you must understand the conditions. The prompt you will send should contain around 100-150 words so as not to be misleading.



Can you provide a short sample that illustrates your “then:” suggestion ?

I would if I could…

First, stuck at “generating image”. Then upon regenerate:


Then a refresh and hung and the pulsing dot.


Try to just ask…




Even “logout” times out…and an oops signing back in.

Everything ChatGPT is borked.

Anyway, here’s custom instructions I was attempting, “how you would like ChatGPT to respond”, dialed back from what I had in there with egregious techniques.

## dalle operation methods
When dalle is being used, these guidelines shall be followed:
- When responding, provide clear confirmation of the user’s request and inform them that the image is being generated. Then do image creation immediately.
- Image generation concludes with a report of success or failure, and the response to user gives the actual json sent to dalle, or the actual error message text received, plus the gen_id that was returned in a success message.

And then the kind of prompt where you can specify just one part is changed:

send to dalle: Subject: The image depicts a colorful macaw parrot’s flight. Style: high quality illustrated fiction. Setting: a rainy and gloomy sky with visible raindrops.

1 Like

Super informative! Thanks a bunch!