Image generation returns unrelated English infographic outputs instead of the requested images

[zattusann3]

What version of the Codex App are you using (From “About Codex” dialog)?

26.623.141536

What subscription do you have?

Pro 5x

What platform is your computer?

MacOS 26.5.1（25F80）

What issue are you seeing?

The built-in image generation path in Codex App is consistently returning unrelated English infographic/template-style images instead of the requested images.

This does not appear to be limited to a complex prompt. After several failures with a detailed anime costume reference-sheet prompt, I tried much simpler isolation prompts. Those also failed and produced unrelated English infographic images.

Examples:

• Prompt asked for an anime illustration of Sakura in a cafe holding tea. Output was an unrelated English "PHOTOSYNTHESIS" infographic.
• Prompt asked for an anime illustration of Yukimasa in a motorcycle suit beside a motorcycle. Output was an unrelated English bedroom-organization infographic.
• Prompt asked for a simple product photo of one Japanese senbei rice cracker on a white plate. Output was an unrelated English image/template, not a rice cracker.

The prompts explicitly included constraints such as "No text", "No labels", "No infographic", "No poster", and "No watermark", but the outputs still contained unrelated English infographic content.

What steps can reproduce the bug?

Feedback ID: 019f356a-451b-76d0-a2fa-bf55201e57ea

What is the expected behavior?

The image generation tool should generate an image matching the user prompt, or fail with an error if generation is unavailable.

For example, when prompted for a single Japanese senbei rice cracker on a white plate with no text, the expected output is a rice cracker image with no text.

Additional information

Reproduction summary:

1. In the affected Codex App thread, ask Codex to generate an image using the built-in image generation tool.
2. Use a simple prompt such as:
   "A simple realistic image of one Japanese rice cracker, senbei, placed alone on a small white plate. Plain neutral background, natural light, close-up product photo style. No text, no labels, no infographic, no poster, no watermark."
3. Observe that the output is not a senbei image and instead appears to be an unrelated English infographic/template image.
4. Repeat with unrelated simple character prompts; the issue still reproduces.

Observed examples:

• Prompt asked for an anime illustration of Sakura in a cafe holding tea. Output was an unrelated English "PHOTOSYNTHESIS" infographic.
• Prompt asked for an anime illustration of Yukimasa in a motorcycle suit beside a motorcycle. Output was an unrelated English bedroom-organization infographic.
• Prompt asked for a simple product photo of one Japanese senbei rice cracker on a white plate. Output was an unrelated English image/template, not a rice cracker.

The issue began after several successful image generations earlier in the same thread. The same thread then started producing unrelated English infographic/template images repeatedly.

This may indicate a routing issue, template mix-up, stale generation context, or another issue in the Codex App image generation path. It does not look like a normal prompt misunderstanding because the failure also occurs with very simple prompts.

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• Severe performance regression on the [$imagegen] skill in Codex App, GPT5.5 need investigation & fixes #31091

Powered by Codex Action

[zattusann3]

Thanks. This does look related.

My case may add a narrower isolation result: the issue reproduces even with a very simple prompt asking for a single senbei rice cracker on a white plate, with "No text". The output was still an unrelated English infographic/template image.

If you prefer consolidating everything into the older issue, I can close this issue and add my reproduction details there.

[zattusann3]

Additional isolation result:

I tested image generation in a different Codex thread, and it succeeded.

Then I returned to the affected thread and tried one more image generation using the successful image as a reference. The prompt asked for the same two adult characters in the same ballroom scene from a different camera angle, and explicitly avoided text, labels, infographics, posters, and English words.

The affected thread still produced an unrelated English educational infographic titled "TYPES OF VERBS".

This suggests the issue may be thread-specific context contamination, routing mix-up, or stale image-generation state, rather than a global failure of my Codex App installation.

Since this now appears to be isolated to a specific thread, I will close this issue for now. I am leaving these reproduction details here in case they help with future investigation.