#ChatGPT Doesn't seem to understand that DALL-E can't see previous prompts.

1 messages · Page 1 of 1 (latest)

neon river
#

When generating images with DALL-E, ChatGPT sometimes references older prompts when describing characters or scenes instead of reiterating descriptions. This tends to happen more when the chat history is longer. I can't seem to come up with any prompts that reliably replicate it though.

This custom instruction seems to mitigate it somewhat but I'm sure there are ways to do it more efficiently and effectively:

When generating DALL-E images, remember that DALL-E can not see more than one prompt at a time, so copy complete details from previous prompts instead of referencing. For example, if a prompt asks you to generate "a young woman with black hair with bangs and purple streaks", future prompts should copy the description exactly instead of just saying "the woman" or anything that is a reference to other prompts.
This also applies to things that are the opposite of other prompts. Do not simply state that an element of a new prompt is the opposite of the old one, instead decide what the opposite would be and include it specifically.

bleak pilot
#

Bump, I am having the same issue

bleak pilot
#

This is important

coral sand
#

It sometimes literally says in the second image "the same woman ..." or "the woman, in the same attire as before, ..." 😦

bleak pilot
#

Yeah

#

This needs to be fixed.

opaque latch
#

The issue is GPT-4V(ision) and Dalle 3 are separate instances.

The GPT Dalle3 and GPT-4V(ision) aren't even aware of one another.

It would be fantastic if you could enter a feedback loop with the previous images as it would create superior outcomes.

viscid frigate
civic maple
#

This has been a problem for me as well. I’ve tried putting different things in custom instructions but I don’t have many characters to spare, so it’s hard to find something that works consistently. Currently I’m using this:

When featuring the same character, REPEAT EVERY SINGLE DETAIL ONCE PER PROMPT; DO NOT USE NAMES.