#Image Gen not successfully prioritizing checking personalization before making image

1 messages · Page 1 of 1 (latest)

vernal lavaBOT
#

Reported by @upbeat imp

Bug Report: Image Gen not successfully prioritizing checking personalization before making image
`Steps to Reproduce`

Have memories or custom instructions on in ChatGPT which give directions related to all image production; Request an image in the chatGPT input, with or without referencing the personalization.

`Expected Result`

Output image will follow personalization instructions as well as context of visible input/output of user and chatGPT

`Actual Result`

Output image only follows context of visible input/output of user and chatGPT, as if contents in personalization are not passed to the image model

`Environment`

Windows 10 Chrome web environment

#
Additional Information

Please provide relevant details to help resolve the issue, such as:

  • ChatGPT Shared Link (if applicable).
  • Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

upbeat imp
#

#images-canvas message

A prompt asking ChatGPT to discuss image preferences succeeds in generating the desired type of image, but even saying:

"Review personalization for image preferences, then create a butterfly image."

Produces the default guess of a monarch butterfly, but asking:

"Review personalization for image preferences, discuss what I like, then create a butterfly image."

Causes ChatGPT to discuss the personalization instructions, putting them in context, and allowing image gen to see and produce a desired image.

Image gen needs to automatically be shown memories + custom instructions that relate to the image request, or chatGPT maybe needs training to input this as part of the prompt to the model.

serene patrol
#

Adding here, previously this workflow was fine when using Dalle. The way I use it: I have a list of keywords that I would like incorporated into the prompt structure for image generation. I’ve worded it in such a way that should be sufficient to instruct the model to build the prompt how I want it to on a first shot basis. Like I said, it was working with Dalle. However, only when specifically instructed after starting a new chat does it incorporate my keywords. The current workaround is to copy paste my instructions each time. Therefore making the use of custom instructions pointless.