Type4_PTS
Well-known member
Ohhhhhhhh, that's interesting! Thx!
.
I believe that what I wrote about that was accurate. But I asked ChatGPT if it was accurate, and this was the response (within the spoiler).
Mostly correct, but one part was a little too absolute.
You were right that another person entering the same 17-word prompt might get a different image. Generative image systems are not photocopiers. Different generations can vary even when the wording is identical, and differences in model version, conversation context, attached images, and other settings can also affect the result.
You were also right that ChatGPT can use saved memories and past-chat information to personalize responses when those options are enabled. OpenAI says Memory can retain useful preferences and interests, while “Reference chat history” may bring relevant information from earlier conversations into a new one. Custom instructions can also influence how ChatGPT interprets a request.
The part I would soften is:
A more technically accurate reply would be:

You were right that another person entering the same 17-word prompt might get a different image. Generative image systems are not photocopiers. Different generations can vary even when the wording is identical, and differences in model version, conversation context, attached images, and other settings can also affect the result.
You were also right that ChatGPT can use saved memories and past-chat information to personalize responses when those options are enabled. OpenAI says Memory can retain useful preferences and interests, while “Reference chat history” may bring relevant information from earlier conversations into a new one. Custom instructions can also influence how ChatGPT interprets a request.
The part I would soften is:
That is reasonable conversational shorthand, but it does not necessarily mean the image generator drew upon every previous prompt you have ever entered. Memory is selective, not a complete permanent transcript of everything. It uses information judged relevant, and some of the richness of that particular picture also came from the surrounding conversation. I had the full satirical passage immediately before the image request, including the office setting, the lawsuit discussion, the couch-jumping reaction, and the exaggerated tone. That context was probably far more influential than the 17 words alone.“ChatGPT knows me well and knows what I like.”
A more technically accurate reply would be:
So HelluvaHoax!’s “ZEN PROMPTOLOGY” compliment was funny, but your micro-prompt had a rather large invisible entourage following behind it.“Someone else using those exact 17 words might get a very different result. Image generation naturally varies, and ChatGPT also had the full conversation for context, including the satirical scene I had described. My personalization settings, saved preferences, and chat history may have helped shape how it interpreted what I wanted, but it wasn’t necessarily generated from those 17 words in isolation.”



