Well, the bug / pattern is visible in normal images - but in this interactive 360 image of a rainforest it’s not visible…
The forest roof looks weird, until you enter the 3d world…
The even bigger problem I see is that this pattern will end up in training data again. And then all generators that have eaten this poison will generate images with these patterns. DallE 3 clearly showed that AIs react sensitively to even slight deviations, and that errors then influence the generators.
That means if the problem is not solved, it will most likely multiply. It will then go beyond this generator.
Yup, those are real crappy images. But you are not showing any prompts - which is the only way to validate that those images were actually created by gpt-image-2.
So, lets see your prompts 
Hi Tina. Just playing around with your prompt:
Prompt
A detailed and colorful image of a a red-tailed hawk flying through dense storm cloulds toward a large emerging rainbow. The bird’s wings must display high-wind turbulence. No land mass shown. The brightness level must be such that all figures are clearly seen.
The prompts are not important and don’t have any role.
This is confirmed by OpenAI itself. It’s a bug in the render engine.
Sorry, your explanation is not good enough.
I really, really, really hope I’m wrong, but there appears to be a F.U.D. operation going on here.
It looks like the model is stumbling when it tries to calculate complex light interactions (caustics) with fine, volumetric, repeating textures. I suspect the neural network is either trying to cut corners to save resources by applying a procedural Voronoi (?) pattern, or it’s getting bogged down in these complex calculations and just outputting whatever is left over. These are just my assumptions.
I tried this prompt, and here’s the result.
Prompt v 1
Scene:
Abstract dark void, dramatic backlit environment, floating embers and glass shards, intense energetic atmosphere.
Subject:
A humanoid form made of molten and shattering obsidian, caught in the exact moment of fragmentation. Sharp, glassy shards suspended in mid-air, catching refractive shafts of cold white light. Deep crimson and orange magma-like glow emanates from the cracks, contrasting violently with the black glass. The pose is dynamic, arms outstretched as if releasing or absorbing energy.
Important details:
High-speed fantasy photography, frozen motion aesthetic, refractive glass textures with caustic light patterns, dramatic rim lighting, particle physics detail (shards, embers, vapor), intense chiaroscuro, surreal energy visualization, masterful composition, digital painting meets photorealism style.
Constraints:
No messy or chaotic composition, keep the central form recognizable. Balance the internal glow with the black glass texture. Preserve the sharpness of the fragments and the fluidity of the molten sections. No decorative borders or text.
AR 16:9
Then I asked Qwen to analyze the image and look for solutions on the web. It cheated a bit by describing slightly larger obsidian shards and removing the dust, but the direct ban on repeating patterns seems to have worked. Either way, the quality of the second image is much higher.
Prompt v2
Scene:
Abstract dark void, dramatic backlit environment, floating embers and suspended glass fragments, intense energetic atmosphere.
Subject:
A humanoid form made of smooth, polished black obsidian, captured in the exact moment of fragmentation. The body is breaking apart into large, distinct, sharp geometric shards of glass, NOT small dust. Internal glowing cracks of crimson light emit energy, but the outer surface of the shards remains perfectly smooth and mirror-like.
Important details:
Macro glass photography, frozen motion aesthetic, individual distinct shards, sharp geometric edges, refractive light patterns, dramatic rim lighting, particle physics detail (large fragments only), intense chiaroscuro, surreal energy visualization, masterful composition.
Constraints:
STRICTLY NO cellular texture, NO webbing, NO neural network patterns, NO repeating Voronoi patterns. The obsidian must look like solid, smooth glass breaking into large pieces, not organic cells or mud cracks. Keep the central form recognizable.
AR 16:9
Posting the second picture in the reply since I’m a new user and can only attach one media per post.
So, you have just proved my point: It’s all in the prompt.
Boyuan Chen @BoyuanChen0
Researcher @OpenAI, core member of the GPT Image research team. PhD @MIT
Issues regarding the noise (spackled/sparkled) bug:
- Thank you for the feedback! I am actively working on it.
- Working on it.
- Will take on it.
- Yes, we’ve noticed this issue, working on a fix now!
- Will fix.
- Etc…
Sorry, still not good enough - just look at above post
If you should be a little interested in the technology: these AI systems do not calculate light or caustics effects like a ray tracer.
In principle, it works in such a way that they analyze an extremely large number of images for patterns and assign them to words. And then from these patterns, they can reconstruct entirely new patterns and so images.
So it does not matter how complex an image would be for a ray tracer, but whether the engine manages to organize the patterns in such a way that they result in an aesthetic image, and follow the prompt as well as possible.
(It is also important whether an image is the first one in a session / after restarting the window, or whether many have already been made. The first images are often a little better than the following ones.)
I will look at the two prompts as soon as I have time and analyze the differences…
What is the difference between the two prompts?
The first prompt causes many small details and microstructures, so exactly what gives the pattern food for the distortion. The second prompt suppresses the complexity and the fine patterns, and causes flat surfaces. With this subject, that works partially. It becomes clearer, but it also loses dynamics and visual complexity.
The decision whether one wants many or few details, high or low fine structures, must be left to the user.
For nature pictures, this optimisation not work!
In the image shown by you, the strategy works very well, it even enhances the image aesthetically and artistically. The black obsidian with smooth surfaces looks more beautiful in large pieces.
But let’s assume you want to create a creature made of smoke in motion, like a tornado that has kicked up many small stones and is swirling them around itself, then it becomes impossible to get a beautiful image.
Anyone who wants to trigger the behavior only has to design the prompts like this:
many microparticles
particle effects
glitter, sparkling
natural rough surfaces
natural textures and structures
fog, smoke, clouds, and movement
nature picture with many small plants
fantastical not existing subjects (Fantasy)
The more fine structure the generator gets, the more it dissolves into clusters.
If you want to suppress the pattern:
flat, uniform artificial surfaces
artificial uniform background
few details
little/no microstructures
no flying particles
no rough surfaces
no natural surfaces or structure patterns
real existing subjects (Photo-realism)
So, in short: leave out everything natural, make it as artificial and flat as possible. As few fine details as possible.
I can also recognize the cluster formation in the second, better image, they appear exactly where higher complexity and patterns can be seen.
A blue sky without clouds is no problem. An artificial environment with flat surfaces and straight lines, simple content, fewer problems.
The opposite triggers the errors.
So, if the goal is to make any naturalness and detail complexity impossible, you would be on the right path. But it is not. The generator should be able to create everything.
The reason… the way a generator works underneath. It is partly a denoiser. In this denoise process, more clusters can form when the structures are detailed and fine than when they have flat surfaces.
The problem is not the prompting, it is the subject!
The point is that the prompt is the path to a workaround
- but you’re apparently fixated on the problem rather than providing tips and the workarounds…
Also I think you’re more mad about people distilling or something…
That’s interesting.
The problem is not the prompting; it is the subject!
Also:
Actually that statement makes zero sense. It’s the prompt that makes the image - or am I missing something?
That’s what I was going to write.
Since this topic focuses on bugs, issues and work-arounds, my personal opinion is that it would be beneficial to share prompts too (obviously optional), just to see if prompting differently makes a different output, like stated above.
I mean I love to see others prompts and learn from it. But subject might also matter, since I’ve seen it myself.
And without ignoring that there seems to be bugs that OpenAI is currently working on.
To answer Jeff: subject is something you choose before the prompt (not always and not everyone that do so), when you have a subject like complex fantasy world, you do the prompting, which if not doing it correctly, can result in unwanted output with ‘noise’. So yeah it’s not dumb idea to share prompts here and let others help to refine it.
Everyone got valid points here (in my opinion). I’m trying to see from every perspective and pattern here.