Codex Image Generation Becomes “Infographic Locked” After a Render Error

[theriver]

Reporting this on behalf of Spirit Studios. Our entire team has been consistently hitting a bug where Codex image generation randomly defaults to producing infographic or poster-style layouts when the prompt has nothing to do with infographics or posters.

It happens across all kinds of prompts and sometimes can happen immediately or sometimes after any number of pictures. So far, there’s no consistent trigger we can identify, but we’ve all experienced it and also have seen others on the internet saying it happens to them too. Starting a new chat clears it temporarily but it almost always comes back. This is really frustrating.

This glitch has been disrupting our production workflow significantly. We’ve not been successful in working around it by declaring output type explicitly at the top of every prompt and adding constraints like “no text overlays, no labels, no infographic layout”.

Requesting this get escalated to the image generation team. Happy to provide screenshots if needed.

— Spirit Studios

[PaulBellow]

Welcome to the community, @theriver !

Do you have any text in quotes in the prompt? Could you share a sample?

You might also add “no infographics” or “no text on image”? Or maybe “An image of…” or “A photo of…” at the start? The reasoning can get confused guessing sometimes.

Hope you stick around. We’ve got a wealth of information and a great group of people here!

[theriver]

Nope there is no text, the prompt will often specify “no text” - but let me show you some screenshots.

What’s worse is that once the image generator makes the error - it is consistent and CANNOT be returned to making normal posts. This is a constant - which is why I know it’s not just a prompt issue. The render sequence MUST be moved into a new chat. One time I used “fork into local” and that actually solved it, but that was inconsistent because it never worked again for not producing infographics immediately.

So, for instance, I recently put in “can you do a pixel based vertical character sheet for all of these characters (reference attached) - this is a full rendered image with transparency. No text anywhere.”

And codex will say: Done. I made a new vertical pixel character sheet with 8 rows and 4 subtle standing/idle pose variants per character, with a true transparent background.

and then i’ll get one or two sheets and then he’ll output this: and never be able to make the paintings again. No matter how many times i prompt - this is what i’ll get.

Often times if i let him continue he’ll say something like:

"The image generator is misfiring and producing unrelated infographic images instead of the Elixir map prompts. I’m stopping the batch here so we don’t waste the whole concept pass on bad generations.

I’ll need to retry once the image tool is behaving normally, starting again with the clean ancient Ethos map."

But then, it will NEVER get fixed… not a day later not a month later. I’ll come back in a week and try another and he’ll say "Yep, that deviation was the image tool going off the rails. Best fix: start a fresh chat with a compact but complete handoff prompt, and make the next assistant load the reference files before generating."

But that said, here are others all over the place discussing. As a new user i can’t share links but you could go to reddit dot com followed by

/r/ChatGPT/comments/1t02at7/sudden_unprompted_image_generation_attempts/

(for one example)

~ ~

Thanks for checking in here and let me know if there’s anything I can do to help solve this!

[PaulBellow]

Hrm. I’m wondering if “memory” is confusing it.

Would there be a way to call separately for the two types of images? That seems kinda clunky, though.

Can you share the prompt that made the Fixed Mindset graphic?

As you said, it’s likely not it, but might be related?

Maybe prepend all requests with either INFOGRAPHIC: or IMAGE: or PAINTING: to see if that clears it up an entire thread? Then, even if it’s in memory, it would know what you’re wanting and not guess?

Sorry I don’t have instant answers!

[theriver]

well that’s the thing i did share one example of a prompt, "“can you do a pixel based vertical character sheet for all of these characters (reference attached) - this is a full rendered image with transparency. No text anywhere.”

and there have been tons of these… Do you want more?

But yes moving forward I can try that and see if it helps in the short term. My hope is that this will get bumped to an actual developer who can look into it and see what might be causing it at a deeper level!

[PaulBellow]

Ah, okay. Sorry. I thought the prompt would be bigger.

@theriver

But yes moving forward I can try that and see if it helps in the short term.

Yes, please let us know.

@theriver

My hope is that this will get bumped to an actual developer who can look into it and see what might be causing it at a deeper level!

Might be best as its own thread then. Do you mind if I split into a new topic?

Then I can try to get some eyes on it as an ongoing issue effecting a lot of people. Thanks!

[theriver]

No please go for it, that would be amazing. Thank you!

I can also provide other examples of prompts if you need

[PaulBellow]

@theriver

No please go for it, that would be amazing. Thank you!

Done! No promises, but I’ve passed it along and will try to get some eyes on it.

Meanwhile, do try the suggestion for pre-pending as I’m kinda curious.

Like, I’m thinking it’s getting stuck thinking you want the same as last image if you don’t specify?

Thanks!

[theriver]

It’s possible… hmm… but often times i have a sequence going like… i’ll upload 20 unique character sheets and i’ll say “make 20 images in sequence of these characters, one at a time, in a unique environment” (i’d specify a lot more but that’s an example) and then it’d break randomly halfway through.

OR I’d give it a script for a video and say “make 10 individual scenes for this script, all unique but with the same art style” and then randomly halfway through or somewhere (or from the start) I’ll get an infographic

[PaulBellow]

@theriver

and then it’d break randomly halfway through.

That’s even weirder. You didn’t suggest once to do an infographic?

I think it’s maybe just guessing tho based on the prompt? Would be great if it were more stable.

I would try some of the prompt tricks I mentioned. Kludgy but it might help?

Let us know! Hopefully someone smarter pops in this thread!

[EricGT]

As @PaulBellow noted about memory, that would be my first ususal suspect.

If you are using ChatGPT versus the API have you tried using an temporay chat? Not that I think it will work, but then again stranger choices have worked at times.

[windysoliloquy]

Is there anything in your agents.md, skills, or docs that might be conflicting with your no-text or infographic request?

Audit your docs with that in mind and see if anything comes out in the wash?

[orangeted]

I am having the same problem. It doesn’t matter what I prompt it’ll start spewing out infographics. Once it starts doing that, infographics is all I get! I have never asked it, at any time, to create infographics or any prompt with that kind of ambiguity. I am using image gen from within codex. I rebooted my machine in desperation and it appears to be fixed. Not sure if this is a clue or not. I’ll come back and let you know if it starts generating infographics again.

Oh… the infographics I get are completely unrelated to the subject I’m attempting to generate an image for. For example I asked it to generate an image of a cardboard cutout of a cartoon plant (with refs) and I got this

[windysoliloquy]

Yeah, that is really disturbing.
This might actually be pointing to an issue with hallucinations being sent into gen.

I’ve only made about 100 images with Codex, but I’ve never seen it do anything close to failing in the way that’s presented in this thread.

Can you ask Codex why it gave you that image when you requested something completely different?

Any clue might help.

[theriver]

Oh thank heavens it’s not just me!!! I feel validated here haha (not that anyone was doubting me, lol)

And yeah to respond to @windysoliloquy - I definitely don’t have any skills running i’m pretty bare bones.

@PaulBellow - I for sure never asked for an info graphic. In fact, it happened again today.

I was running a series of graphics for an educational program i’m making about Tarot, actually. 22 images for the Major Arcana - and then around scene 17 it did this: Interestingly, I got a dog this time too, so that’s… a nice mix up You can see my prompt though (this was actually a steer because he got one wrong, but steer or not this still happens), where I said “continue making images with render engine and no text anywhere”

I will say though, prompting since yesterday morning “make a painted image with render engine and no text” DID seem to help, or i just got lucky. I hadn’t had any more of these glitches until now, I thought it was somewhat solved! Apparently not, but it was a bit better.

Regarding Memory, I put this into the memory about 4 days ago and it still happens, is this what you meant? This is all that’s in there right now:

Always use the image generation workflow when asked to create an image

Never create infographics at all

[windysoliloquy]

You run no skills and are bare bones.

I had a really bad image generation flow with some nastiness that got written into a doc by codex… small error but it sat in the repo docs.

So we have two different fail modes, but both have a ‘locked’ common thread…

That’s quite a little puzzle.

I had a persistent error that I found in the docs codex wrote, by having codex audit the docs…

But the error was nothing like what you guys are showing.

[theriver]

Honestly, I was very surprised that there weren’t more people talking about this, because when I coordinated with my team to start generating scenes for this course that we’re making, anyone using Codex in the Codex app was getting this problem.

I’m really hoping we can get this seen by some developers who might be able to figure it out because it’s very strange!

[EricGT]

@EricGT

tried using an temporay chat?

ChatGPT replied

I can’t generate images in this temporary chat because the image-generation tool isn’t available here.

[LarisaHaster]

Hi @theriver,

I’ve been following this a bit under the shadows and I also did a test run (GPT-5.5 medium, standard) and couldn’t reproduce this, although I understand this issue may be intermittent and appears more often during longer image sequences.

While reading this topic, I noticed one thing I have not seen clearly compared yet is whether it happens across every GPT-model/intelligence/speed setting or only under certain settings. Now I don’t know for sure if this might be of relevance.

But maybe it may also help to compare:

• a single standalone image request versus a long sequence
• the first image in a fresh thread versus later generations
• whether changing the Codex setting after the infographic behavior begins makes any difference
• whether the same prompt works correctly when copied into a new thread
• whether the failure tends to appear after repeated “continue,” steer corrections or many uploaded references.

Since GPT-image-2 is the image model underneath regardless, these comparisons might help show whether the failure is coming from the image model itself, the GPT model’s image-tool orchestration/tool call or accumulated thread state.

Hopefully I am not repeating what you’ve already told here (or misunderstanding), otherwise apologies before hand.

[orangeted]

This after producing many ‘good’ images (so I know the workflow works)