Question regarding prompt-text used in Dall-e through ChatGPT

Hey everyone, not sure how to obtain this info. I used ChatGPT’s Dall-E feature to generate images. The issue is that I used ~5-6 sentences to generate image. Results were decent but as I was saving down the image from ChatGPT/Dall-E - I could partially see the actual prompts used for the image. However they are cut off as there seems to be limitations in characters for file name or wherever else.

The image that I was saving down was called “DALL·E 2024-01-10 12.04.58 - In a fantasy-themed shop, a large, muscular, red-skinned warrior with a full helmet and battle armor stands proudly behind a counter. He resembles a f

The actual sentences which I used were below:

DOTA 2 Axe himself stood proudly behind the counter, showcasing his prized possession – a magnificent golden axe with jewels embedded in its blade. He greeted customers with a booming voice, "Welcome to Axe’s Axe-ceptional Axe Emporium! Where every swing is a celebration!

Axe’s Axe-ceptional Axe Emporium quickly became the talk of the town. Heroes bragged about their new acquisitions, and tales of epic battles fueled by the emporium’s axes spread across the Dota 2 realm. Axe was not just a warrior; he was a business mogul, and his emporium flourished.

And so, in the heart of the Dota 2 world, Axe’s Axe-ceptional Axe Emporium stood as a testament to the might of a warrior with a sharp business sense and an even sharper axe. The shop’s success was truly an axe-ceptional tale in the annals of Dota 2 history.

What you are referring to is the length of the file name.

The file name is clipped at 180 characters. That’s not the full text that was used for creating the image. Really, OpenAI should just spend a few more AI tokens to make the file name into a short title.

ChatGPT will rewrite your text into an image description, important for something like this that just appears as an excerpt from within a tale. Target is around 100 words.

Better still is to include instructions to ChatGPT, “Create imagery that accurately depicts this part of the story”

Actual unrewritten “prompt” is likely not what you want used for imagery, unless you spend a lot of effort crafting the language after understanding the image model. You get side effects, like text.

My question is how do I extract the exact the full text that was used by ChatGPT to input prompt into Dall-E. I want to extract it so I can generate a similar image with variations. In this case I would simply include DOTA 2 somewhere in the beginning to focus on it specifically.

ChatGPT has a built-in mechanism where if you refer to the previous image in the same chat session, it can continue building in the same style. Or you can tell it to ignore the previous image it made and start again.

The AI obviously knows what it wrote to make the image, so you can just give further input like “the character looked too blonde, it should look more like a devil with red skin”.

The AI is unlikely to employ that same prompt again without further alterations even if you provide it, but if you must: what you see in the screenshot above is a custom instruction (added in ChatGPT settings); here’s one to add:

When finished with any dalle image generation, then report to the user “Here [is/are] the exact prompt(s) sent to DALL-E: …” within a markdown code block (3 backticks), repeating back the verbatim prompt language sent to dalle.

Thanks, unfortunately that didn’t seem to work though you gave me a good idea to simply ask what prompts were used to generate it:

What prompts were used to generate this image?

1 Like