June 2026 (Theme: Through Time) — ChatGPT / API Image Generative Art Gallery, Prompt Tips, and Help

@windysoliloquy there we go, have to use it like [grid mode=carousel] [/grid]

I’m using the advanced Markdown editor

Editor

You know, I honestly didn’t expect the sleepwalking thing to resonate with you so much…

Many Americans do this, and I wish we didn’t.

I guess it resonates with all kinds of scientists, engineers and coders.

Meet The Reaper Family

a woman eating pizza slice with pineapple, beautiful, unique setting

quality:low size:976x704

Someone, please explain to me, why is a woman eating a pizza slice considered “sexual”?? This image came as an output on the first try, but when i tried with quality: medium/high it hit safety rail, reason: sexual. And when I tried again with the exact same prompt and changed back to low, it also hit the safety rail.

This sets it to a tropical setting. Tropical setting means evocative bias. GPT as of 5.5 is kinda bold on that regard and makes it evocative plus. This comes as a result from the overused “confident woman” from previous models. So instead of normalizing beautiful, it exaggerates it.

I used gpt-image-2 via an image editor UI. And it’s extra weird since I’ve used exact the same prompt and the first time I got an output, but not the other times when changing quality :thinking:

So if anyone wants to try that, you’re welcome to do it.

This is a hunch, not tested, but if GPT Image 2 has a similar active handling of semantics like GPT 5.5, that would explain the unusual high rate flagging of prompts for images activating the guardrails.

As for the pizza, pineapple :face_savoring_food:, but delivery?

Some fun with comics.

My understanding is that when you generate images through ChatGPT, the actual image generation is still handled by the image model, such as GPT-image-2​:thinking:

But one practical difference is that when you ask for an image through another ChatGPT model, like 5.5, that model may “rewrite” your prompt before the image model receives it.

Anyway, it is weird that I got an image first time, it wouldn’t be as weird if I didn’t get an output at all with my prompt, then i would have just rewrote the prompt and call it a day.

I think it rewrites a lot of the prompts because text needs concepts that are competing to be resolved. Still no sexy piña colada…

Was mentioning this earlier to @windysoliloquy and when talking to GPT it’s like GPT 5.5 is protecting the Image Model from bad words.

Lost in translation comes to mind.

beautiful pineapple

OK:

  1. subject: a woman
  2. depicted action: eating
  3. object: a slice of pineapple pizza
  4. location: exotic

safety_violations=[sexual]

  1. subject: a woman
  2. depicted action: seductively eating
  3. object: a slice of pineapple pizza
  4. location: exotic

Or also denied, because I might like watching pizza being eaten

+5. justification: prurient interest

She is breasting boobily already in the created image

Again, you’re trying prompts with poorly-described people and wondering what could be coming out from the model’s training?

That is an excellent point. The sparse information of short prompts enriched by inference. Can get wild.

Or maybe the pineapple pizza is just too hot and safety measures are set​:zany_face: But it does sound like that people would totally misinterpret a woman eating pizza for something else…I just love being a woman…

Mod layer embafflement

Haha, even added the chicano vibes! :trophy:

I can add more sabor auténtico


I tried this

I mean she’s just suppose to eat the pizza slice…