OK hurrah! Bug fixed in few days, thanks to the tech team!
Noise pattern amplification is gone, and every new generation is a new picture.
(I think the noise is still there, but it does not amplify anymore.)
My first impression is, the new 2.0 is REALLY better than 1.0…
And it is recommended now to send an exact pixel ratio like 1536x1024, landscape can be different ratios now.
Definitely not the case for me. I’m still seeing blatant artifacting and noise spawned from reiteration of previously generated images if requesting any further gens in the same chat. Here is a fresh example from my results.
Hello, welcome in the forum.
Do you have re-upload the images?
The noise is still there, and it can still amplify if you reuse pictures.
It just does not amplify in every new generation anymore.
If you generate some pictures in the same session with the same prompt, you should not get the amplification effect anymore. This is fixed.
But reuse will still show the same effect. You can demonstrate it if you want, or i will do it during testing time.
Hello.
I didn’t re upload any images, or use any reference pictures for the generations. I just prompted images in the same chat, but saw continually worsening noise with further iterations, like I was before.
If you have problem with image generation 2.0 in ChatGPT, Boyuan Chen can help.
Best for him would be to check the forum too, i think.
OK, interesting… not on my account. It does not amplify anymore. But yes, the noise is there and it disturbs me.
I think we must wait a few days to see what happens, I see they are working on fixing it.
it could just be that the fixes are still rolling out and I haven’t gotten it yet. Regardless, I am certain that they will fix it at some point, as it’s an issue that many people are noticing and it’s hard to miss.
Can you please supply your prompt so we could take a look? Look like your size resolution was 1536x2048 (non-standard).
that’s just a merging of two separate generations I placed in photoshop because I couldn’t post two separate images due to my forum account being too young.
Welcome to the Community
Once you get your account going, we look forward to your participation. In sddition to this thread, feel free to check out this one:
It is definitely NOT fixed. Only the image carryover has been corrected.
How strong the pattern is depends on the motive.
(Guys/Developers… don’t put the stuff into the latency space…)
Dense fantastical jungle, oversized plants with spiral-shaped trunks and broad shimmering leaves, glowing fruits hang like small lamps, between the roots flows a calm river of dark water with reflective sheen, delicate vines form natural arches, large alien creatures with soft skin move slowly in the background, ground shows moist organic textures in rich green and deep brown, warm diffused light falls through dense canopy and creates vivid patches of light, high humidity visible through fine haze, lush, vibrant, mysterious, highly detailed, immersive depth. Photo style.
It is definitely NOT fixed. Only the image carryover has been corrected.
How strong the pattern is depends on the motive.Guys… don’t puts that stuff into the latency space…
It would be helpfull if you showed your prompt. And I don’t quite understand what you stated.
I not prompt in english. So not sure if it is the prompt the system used.
OpenAI has published a Cookbook entry for gpt-image-2 aimed at API users. The prompting strategies in this guide are also applicable when using the same model through ChatGPT.
This guide highlights prompting patterns, best practices, and example prompts drawn from real production use cases for
gpt-image-2. It is our most capable image model, with stronger image quality, improved editing performance, and broader support for production workflows. Thelowquality setting is especially strong for latency-sensitive use cases, whilemediumandhighremain good fits when maximum fidelity matters.
Release date: April 21, 2026
that’s valuable too…
YOINKED and popped into the gallery thread
Yeah. Not fixed. Same issue.
Prompt:
A breathtaking, photorealistic fantasy scene portrays a majestic, golden-brown dragon with outstretched wings perched atop a jagged, rocky cliff, its sharp purple eyes gleaming. The surrounding landscape features vast, misty mountains and a soft, cloudy sky, emphasizing the dragon’s imposing presence in an awe-inspiring, cinematic setting.
It’s the 7th prompt in a series of trying to get it right.
Same issue with the “checkerboard pattern” here (don’t know how to name it).
I create images using GPT since about 3 years now and the quality has never been this worse (in technical terms speaking).
This image (above) is the 1st in my conversation, set by a prompt (not a reference image). And you can already see the “square pattern / structure / texture” in the artefact and soil (ground).
When I alter the prompt, it becomes worse (2nd generation, not altered by the “edit” button, but by modifying the prompt itself).
This was the 3rd attempt: all of them are not usable for professional usage.
Same here: it’s not smooth, but the layering are made out of chunks. Or something: very ‘blocky’.
And also in non-realistic images: you can clearly spot the blocks/ square pattern in this graphic (fresh creation, not altered).
Another problem is the LLM itself: it doesn’t listen to my image workflow at all.
For more than three years I had my agents set-up like this;
But since Image model 2.0 it goes like;
Sometimes it even creates an image without any subject! It immediately starts creating an image, not following my workflow.
I created about 25 iterations of it, and it never summarizes the prompt before creating, or is asking for my confirmation.
Worst of all is that sometimes it creates an image of my workflow in another language. Like it’s repeating it’s instructions and styles, but as a text-output (generated in an image).
Sorry to say, but this new model is not usable for serious usage.
An idea for the developers: If this stuff damages the weight data of other AI systems, if they end up in training, it will act like microplastics in the data-sphere. And just like plastic manufacturers, data manufacturers could then be sued as well?
It is the wrong way.
You mean: Droste effect - Wikipedia
