Generating Random Non-Related Image While Using Codex
Open 💬 5 comments Opened Jun 20, 2026 by HENGRUIZZZZ
What version of the Codex App are you using (From “About Codex” dialog)?
26.616.32156
What subscription do you have?
Pro
What platform is your computer?
MacOS 26.2 (25C56)
What issue are you seeing?
While building a health-related website, I used the ImageGen tool to generate icons for the interface. However, instead of creating the requested health-related icons, it generated a series of unrelated images about “Boolean Operators.” In addition, it unexpectedly surfaced memories or references to “Timeline of Human History” and “The Water Cycle”—topics that I have never previously generated, requested, or worked on. These outputs appeared completely unrelated to my current task and project context, suggesting that the tool may have retrieved incorrect context or generated content from an unrelated source.
What steps can reproduce the bug?
Using GPT5.5 Super High, Fast Mode
What is the expected behavior?
_No response_
Additional information
_No response_
5 Comments
I'm having this issue too, codex is getting very confused about why the image tool is only spitting out random infographics instead of the icons we've been making for the last hour
<img width="1668" height="1298" alt="Image" src="https://github.com/user-attachments/assets/276f588a-cb1c-498a-b289-386d3d34ed00" />
+1
I've encountered the same bug for the past week, so I investigated.
When I use ChatGPT image gen in the browser, it follows the reference images way better: mascot identity, pose, composition, etc. But when I use Codex’s built-in
image_gen, the result feels much less controllable and often drifts pretty far from the same task.The confusing part is that the
$imagegenskill makes it sound like local/reference images can be used if they’re visible in the conversation first. But the actual tool surface I can see in Codex only exposes apromptfield. I don’t see anything for image inputs, file IDs, reference images, input fidelity, edit mode, quality, etc.Also, the Codex pricing page lists separate GPT-Image-2 rates for
GPT-Image-2 (image)andGPT-Image-2 (text):https://developers.openai.com/codex/pricing
So it seems like Codex is accounting for image-side usage somehow, but the exposed tool doesn’t give any explicit way to pass/control image references. That makes it hard to tell whether references are actually being sent to the image model, or whether they’re just being summarized through conversation context.
from dmeese:
me too,this is a terrible super BUG;