Severe performance regression on the [$imagegen] skill in Codex App, GPT5.5 need investigation & fixes
What version of the Codex App are you using (From “About Codex” dialog)?
26.623.101652
What subscription do you have?
PRO
What platform is your computer?
MAC M4PRO Darwin 25.5.0 arm64 arm
What issue are you seeing?
<img width="720" height="420" alt="Image" src="https://github.com/user-attachments/assets/cd7d84ae-cb31-49ab-a14e-bb82824d0936" />
<img width="840" height="510" alt="Image" src="https://github.com/user-attachments/assets/dc7a2f80-497d-4321-9758-b257a41107f6" />
<img width="583" height="247" alt="Image" src="https://github.com/user-attachments/assets/dc0eadeb-0f0e-45cd-9bb4-ed303bc5a3b9" />
What steps can reproduce the bug?
The [$imagegen] skill in my Codex App has gotten way worse performance-wise lately, and I hope the team can look into this bug and push out fixes.
App version I'm running: 26.623.101652
I prompted it to make 3D-style icons, but the outputs constantly go against my prompts. All the icons it cranked out are plain, minimal flat designs instead. I've attached an icon I generated using the exact same [$imagegen] skill back in early June, and the gap in output quality is really night-and-day.
I've spotted another weird flaw with the [$imagegen] skill powered by GPT-5.5 on this 26.623.101652 build: it often deliberately veers away from prompt instructions and spits out totally unrelated imagery. I've attached two such irrelevant pictures as bug proof — the tool itself claimed those outputs were supposed to be icons. After it realized those results missed the mark, it only gave a short note saying the design had gone wrong. It felt like the underlying model got downgraded right after that; the icons it produced afterwards turned out even lower-quality and stuck to the unwanted flat aesthetic that broke my 3D requirements. Worst of all, the app mistakenly insisted those flat icons actually matched my requested 3D look.
I ran into all these issues when working with Codex App throughout late June to early July. For comparison, I tested Codex CLI using the same prompts, and it never showed any of these glitches, so the problem seems exclusive to the desktop app for now. I hold the Pro subscription, so I definitely didn't expect to hit such obvious drops in image-generation quality plus frequent prompt-drifting issues on GPT-5.5's [$imagegen] skill.
What is the expected behavior?
_No response_
Additional information
_No response_
2 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
The test results are crystal clear: the built-in Image Gen pipeline is glitching out in this session.
I used the default built-in image_gen mode exactly as the skill specs laid out, and I didn’t pull any pre-made web icons at all. I ran two separate generation attempts:
New 3D icon for a vendor stall permit
A trimmed-down prompt to generate just one standalone 3D game icon
Both outputs were totally off-track:
The first one spit out an AI evolution infographic instead.
The second turned into a liver health fact sheet titled Liver - The Body's Detox Hero.
It completely ignored hard restrictions I added like "single mobile game icon, no text, no infographics, no titles, no labels" both times.
This rules out the possibility that I swapped in ready-made web assets by mistake—this was a genuine call to the native image_gen tool. The outputs having zero relevance to the prompt points to a routing bug or cross-contaminated context within the built-in image_gen tool, not just poorly refined prompt wording.
I won’t save either of these broken outputs to the project folder, nor will I swap them into the app’s asset library.
We’ve got two viable next steps to pick from:
Option 1: Keep testing the native Image Gen
Run one more test with an ultra-simplified prompt, something super basic like: a single 3D blue-and-gold game icon, no text.
If it still generates unrelated graphics, we can confirm the built-in tool is unreliable for this round’s icon asset creation work.
Option 2: Switch to the CLI fallback or local design workflow
If you still want fully original AI-generated artwork: Go with the CLI fallback specified in the skill docs, which requires a local OPENAI_API_KEY.
If consistent, production-ready app assets are your top priority: Use existing web-based 3D assets as a style reference, then do controlled touch-ups and compositing via Pillow partial inpainting. This counts as engineering-level asset retouching/compositing, not raw native generation from image_gen.
Wrap-up takeaway for this round
This isn’t misaligned requirements on your end, nor did I cut corners by reusing stock assets. The built-in image_gen tool’s outputs bear almost no relation to input prompts right now. It’s not stable enough to plug straight into our app icon production pipeline for the time being.
<img width="1672" height="941" alt="Image" src="https://github.com/user-attachments/assets/325bd03c-068c-4d97-a634-11836e8ee6ac" />
<img width="1536" height="1024" alt="Image" src="https://github.com/user-attachments/assets/2820df99-da60-446f-81dc-6e72bd090988" />