OpenAI's GPT Image 2 arrived on April 21, 2026 to well-deserved fanfare. The model brought with it a genuinely rebuilt architecture — no longer running on the GPT-4o image pipeline that powered its predecessors, it introduced native reasoning ("thinking mode") into image generation for the first time, letting the model plan layouts, pull web references, and self-check outputs before delivering a result. Text rendering jumped to 99% accuracy in English and crossed 90% in Chinese, Japanese, Korean, Hindi, Bengali, and Arabic. Aspect ratios expanded dramatically, from a tight three-option list all the way to anything between 3:1 and 1:3, including native 16:9. Standard output sits at 2K resolution, with 4K available in API beta. Within hours of launch, GPT Image 2 claimed the top spot across every category on the Image Arena leaderboards, with a text-to-image ELO score sitting more than 240 points clear of the next closest model. By most measures, it was a remarkable debut.
But almost from the first generation, a growing number of users noticed something wrong. Scattered across images — textures, skies, backgrounds, fabric folds, even smooth architectural surfaces — was an uninvited guest: a faint but persistent pattern of tiling grime, diagonal grids, or wavy digital noise that had no business being there. The kind of artifact that makes an otherwise stunning image feel subtly, deeply broken. This is the story of what that artifact is, where it might come from, and what you can actually do about it.
GPT Image 2's Tiling Texture / Grime Artifact Problem
The artifact goes by several names in community discussions — "tiling texture," "grime," "the noise," "checkerboard pattern," "digital ripples," "blocky overlays" — but they all describe roughly the same phenomenon: a repeating, structured visual contamination baked into the image at the pixel level, often most visible on surfaces with natural regularity. Clouds, stone floors, fabric weaves, wooden textures, skin, grass, and water are all especially susceptible. In images with cleaner subjects, the artifact is subtler but still detectable to a trained eye — the kind of thing that feels like compression artifacting from a badly encoded JPEG, or the "blubber distortion" of a heavily compressed MP3 file, but rendered in visual form.
One early documenter on the OpenAI developer community forum described the sensation precisely: it causes a kind of neurological stress, the same way lossy audio artifacts do, even before the pattern becomes outright destructive.
The problem reveals itself in two related but distinct forms. The first is a baseline noise present in fresh generations — even a first image, in a clean session, with a straightforward prompt, can carry the pattern. One user generated a scene of an elderly dockworker on a rusted anchor — a prompt with detailed, careful lighting specifications and explicit instructions against heavy grain — and still got "digital ripples" washing over the lower portion of the frame, blurring the character's boots and legs. Another generated a fantastical jungle scene and found the ground textures visibly chunky and grid-like.
The second, more severe form is noise amplification: when a user continues generating images in the same chat session, the artifact compounds. Each successive generation inherits and intensifies the noise from the previous one. By the third to fifth image in a session, the results can become genuinely unusable — the same corrupted pattern stamped across entirely different subjects and styles. Users who tested this systematically — generating five images of the same prompt in sequence — documented the degradation in almost identical language: the images become more and more identical as the noise takes over, and the generation loses creative variety entirely.
The amplification bug proved to be the more dramatic complaint, and OpenAI's engineering team addressed it within a few days of launch. But the underlying baseline noise — the faint tile pattern present even in clean first generations — remained. As one forum member put it plainly after the patch: "It is definitely NOT fixed. Only the image carryover has been corrected."
It is worth pausing here to address a question that frequently comes up: if the artifact is not crashing anything or throwing an error code, why is it such a big deal? Daller, the forum thread's primary maintainer and one of its most technically precise contributors, offered a useful framing. The artifact follows a consistent rule: it intensifies wherever there is high visual complexity — glass structures, rock surfaces, structured clouds, sand, and anything with fine repeating microstructure. In flat, uniform areas of an image, it is much less visible. The effect behaves, in his words, like an acoustic hum heard underneath music — for someone with sharp pitch sensitivity, it is immediately noticeable and genuinely disruptive; for most others, it only registers once it becomes dominant enough to distort the overall image. He was also careful to distinguish this from DALL-E 3's characteristic flaws, which were largely training data artifacts. The GPT Image 2 problem, by his assessment, is something different: not a training data quirk, but a wrong technical process in the generator itself. That distinction matters, because it implies the fix requires a change to the generation pipeline, not a data re-run.------
Why Does This Happen?
OpenAI has not issued a formal public explanation for the artifact. There is no official post, no model card note, and no technical document addressing it directly. What exists is a rich thread of community speculation, some of it quite technically sophisticated.
The Steganographic Watermark Hypothesis
The most debated theory is that the tiling pattern is not a bug at all, but an intentional feature: a steganographic watermark embedded into every generated image to allow provenance tracking and AI-detection. DALL-E 3 already used invisible watermarks detectable by OpenAI's own tools, and GPT Image 2, with its architectural rebuild, may simply have pushed this practice deeper into the generation process itself. The pattern's consistency across wildly different subjects and styles lends some credibility to this reading — if it were a pure artifact of the generation process, you would expect it to vary more with image content. Its regularity, by contrast, looks more like something that was added on top.
A community member who first surfaced this hypothesis noted that it would also explain why the artifact is robust to prompt-level suppression: you cannot instruct the model not to embed a watermark, any more than you can instruct a printer not to print the yellow microdot pattern required by its manufacturer.
Latent Space Interference
A second school of thought points to something going wrong in how the model constructs detail in the latent space during generation. GPT Image 2 uses a hybrid architecture that, by OpenAI's own admission, has not been fully publicly explained. Several users with backgrounds in image generation have speculated that the artifact could result from an incorrect method of adding fine detail — specifically, from manipulating the initial noise or latent state in ways that introduce structured patterns before the model has a chance to render them away. The user going by the handle Daller put it bluntly in a message to the developers: "Guys, don't put the stuff into the latent space."
Depth-First Composition and Structural Seams
Another observation pointing in a related direction came from a close reading of the model's compositional behavior. When examining the black background of a coffee machine image, one user noticed a grid of layered, interconnected rectangles — each precisely shaped to contain a distinct visual element: pipes, valves, mechanical components. Viewed as evidence of a depth-first construction process, this suggests the model defines structured spatial regions first, then populates them — and the seams of that underlying grid may be what bleeds through as the artifact in complex scenes.
The Denoiser Cluster Hypothesis
Forum user Chain_L proposed a more mechanism-focused explanation: the model appears to struggle specifically when handling complex light interactions — caustics, refractive surfaces, fine volumetric textures — and responds by either applying a procedural Voronoi-like pattern as a shortcut, or simply outputting whatever the computation produced when it ran out of capacity. He tested this directly by generating an obsidian humanoid shattering into glass shards — exactly the kind of high-caustic, high-detail scenario that would stress a light-calculation system — and observed the characteristic artifact across the refractive surfaces.
In a follow-up post, he revised the prompt with explicit exclusions against Voronoi and cellular patterns and found the second generation significantly cleaner.
Daller responded with an important technical correction: the model is not actually computing caustics like a ray tracer at all. It is a denoiser at its core — it analyzes patterns in training images, maps them to concepts, and reconstructs new patterns to match a prompt. The real question is not whether caustic calculations are too expensive, but whether the denoiser can organize the patterns coherently when the subject contains extreme fine-grained detail.
His conclusion: the more microstructure a prompt demands, the more the denoiser dissolves into clusters during its reconstruction pass. He went so far as to compile two lists — one of prompt elements that reliably trigger the artifact (many microparticles, particle effects, glitter, natural rough surfaces, fog, smoke, fantastical subjects), and one that tend to suppress it (flat artificial surfaces, uniform backgrounds, few details, photorealistic real-world subjects). "So, in short," he wrote: "leave out everything natural, make it as artificial and flat as possible." He acknowledged the obvious problem with this as a solution: the whole point of the generator is to be able to render everything.
Context Contamination from Web-Referenced Style Lookups
A more targeted observation came from a user going by the handle summerstay who noticed that the artifacts were consistently worse when the model was asked to generate in the style of a specific artist or movement — particularly when the model's thinking mode had gone out to search the web for visual references. The theory here is that the retrieved reference images, when fed back into the generation context, introduce a kind of visual noise that contaminates the output. This would explain why the problem is more pronounced with "in the style of" prompts and in sessions where reference images have been uploaded. The same user found that prompts with no reference images and no style lookups often produced significantly cleaner results.
Later, summerstay provided a concrete example to back this up. They had prompted the model to paint Joan of Arc in the style of Alphonse Mucha's Slav Epic— and watched, through the model's thinking notes, as it went out and fetched reference images of the original work. The resulting image is striking in many respects, but look closely at Joan's outstretched arm, and you can spot exactly the kind of light and dark splotches that characterize the artifact elsewhere in the thread. The reference lookup had left fingerprints. The same user also observed that prompts with no style references and no uploaded images — clean, purely descriptive text prompts in a fresh session — could produce outputs with no visible artifacts at all.
Unrecognized Art Styles as an Artifact Trigger — or a Prompting Problem?
The style lookup problem takes on a different character when the style in question does not exist at all. When a user names a style the model has no training coverage for, it has nothing to retrieve and nothing to fall back on — and the question becomes whether the model handles that gracefully, or whether the gap itself becomes a source of degradation. The community found a case that puts this question in sharp relief.
Taragonn prompted the model with a detailed atmospheric scene — a frost-covered road, a fort on a hill, fallen soldiers, oppressive dawn mood — specifying it be rendered in "fantasy art in Narthmor style." The output came back artifact-heavy. His post also included an interesting piece of evidence: the internal prompt ChatGPT had actually sent to the image generator contained no mention of "Narthmor" at all. The model had silently substituted the unrecognized style name with its own descriptive English translation before passing the request on.
windysoliloquy offered a counterargument: "Narthmor style" is not a real recognized art style, and the degraded output may not be a model bug at all. He then ran a follow-up test — stripping "Narthmor style" from the prompt entirely while keeping the rest of the scene description unchanged. The result was noticeably cleaner. His interpretation: the unrecognized term carries significant weight in the model's processing, and forcing it to resolve something it has no reference for degrades the output. Remove the unknown variable, and the model generates more cleanly. Whether this confirms a prompting issue or simply demonstrates that unknown style names trigger the same underlying artifact seen elsewhere is, of course, a matter of interpretation.
Training Data Contamination — A Compounding Long-Term Risk
User iyvljtjqleyj raised a concern that extends beyond the immediate problem: if images with the characteristic blocky, pattern-heavy artifact are published on the internet — which they already are, in large numbers — future models trained on web-scraped data will encounter them and learn to treat them as valid visual targets. The result could be a self-reinforcing cycle in which the artifact becomes progressively more baked into generation behavior across model generations, not less. They drew a historical parallel: Image Generation 1.0 was trained heavily on Instagram images with warm yellow-movie-style filters, and for a long time produced images with a pervasive yellowish tint as a result. The same mechanism that caused that problem could, if the artifact persists long enough in the wild, cause a similar drift in GPT Image 2's successors.
Wave Interference in the Generation Space
One more physically-minded hypothesis compared the artifact pattern to optical wave interference — specifically diffraction rather than refraction. In this reading, as the model generates fine detail, frequency components in the generation process overlap and interfere: in some regions the waves reinforce each other, producing intense clusters; in others they cancel out, producing voids. The result is the alternating bright-and-dark splotch pattern visible especially in cloud swirls and feather microstructures.
A user who tested a red kite in flight — a prompt specifically designed to capture turbulent airflow and fine feather physics — found that the model smoothed out the bird's defining forked tail feature entirely, while the cloud background showed exactly the interference pattern described.
How to Work Around the GPT Image 2 Tiling Texture Issue
Until OpenAI ships a complete fix, the community has collectively assembled a small toolkit of partial mitigations. None of them fully eliminate the artifact, but they can meaningfully reduce its severity.
Start Fresh Sessions and Avoid Reusing Generated Images
The most reliable advice — and the most inconvenient — is also the most consistent across every discussion thread: do not continue generating images in the same chat session once noise has appeared. This is not merely a ChatGPT interface problem. Forum user _j tested the API edits endpoint directly — iterating on the same castle scene across multiple calls, with each output fed back in as the next input — and confirmed that after just a few passes, the same mottled clouds and blotchy overlay patterns appear in the API output as they do in ChatGPT. The noise accumulates wherever image input is involved, regardless of whether you are in a chat session or making stateless API calls.
Starting a new session with a text-only prompt clears the accumulated context and gives the model a cleaner baseline. Similarly, if you need to use an image as a reference for further editing, consider extracting a prompt description from it first and using that prompt in a fresh session rather than feeding the image back directly. One of the forum's earliest documenters tested exactly this — generating a prompt from an uploaded image and using that prompt to re-generate — and it produced noticeably cleaner results than direct image re-use. (We have written a more detailed breakdown of this specific ghosting behavior separately, for those dealing primarily with that variant of the problem: GPT Image 2 Artifacting: Previous Chat Images Leave Traces in New Ones.)
Follow Up With a Cleanup Prompt
For images where the noise is visible but not catastrophic, several users have found success asking the model to clean up its own output with a follow-up prompt along the lines of "Remove the noise from the image while keeping all the lines." This is not a guaranteed fix, and it works better on some image types than others, but as a lightweight first step before abandoning a generation entirely, it is worth trying.
Explicitly Exclude Artifact Patterns in Your Prompt
As discussed earlier, explicitly telling the model what not to render — "STRICTLY NO cellular texture, NO webbing, NO neural network patterns, NO repeating Voronoi patterns" — can produce meaningfully cleaner results. The complementary principle is to describe surfaces in terms of what they should look like rather than leaving the model to infer texture from a style description. The tradeoff is real: suppressing the artifact this way also suppresses visual richness, so this approach works best for subjects that tolerate clean, geometric aesthetics.
Avoid Style Reference Prompts — Especially Unrecognized Styles
When using "in the style of" prompts that trigger the model's web-search-backed reference lookup, expect more visible artifacts. If a particular style is essential to your workflow, you may find better results by describing the style in your own terms — lighting conditions, color palette, compositional approach — rather than naming an artist or movement directly. This reduces the likelihood of the model pulling external reference images into the generation context.
More critically: avoid referencing styles the model cannot confidently recognize. A fictional or highly obscure style name does not cause the model to default gracefully — it appears to be one of the conditions most reliably associated with severe artifact output. If the model does not know what a style looks like, it cannot generate cleanly in that style's absence.
Consider the API for Text-Only Input Workflows
For workflows that rely exclusively on text-to-image generation — no uploaded reference images, no iterative edits using previous outputs — calling gpt-image-2 via the API's v1/images/generations endpoint avoids the session-state accumulation problem that plagues ChatGPT conversations. The artifact can still appear in baseline generations, but the compounding amplification effect is not present when no image context is being passed between calls.
If you want to experiment with this path without navigating OpenAI's tier requirements, APIPASS already has GPT Image 2 API available. It is a reasonable starting point for testing whether stateless text-only generation produces meaningfully cleaner results for your specific use cases before committing to a full integration.
When Will OpenAI Fix the Problem?
On April 24, a user on X named @RifeWithKaiju asked the question that was on a lot of people's minds: has there been any official word on the artifacting? They described what many others had also noticed — that the artifact does not look like one thing, but many things depending on the image. On some pictures it looks like pointillism, on others like JPEG compression, on others still like multi-colored stars scattered across the entire frame. The noise amplification bug had already been somewhat alleviated by then, they noted, but the underlying artifact was very much still present.
Boyuan Chen — an OpenAI developer directly involved with GPT Image 2 — replied the same day with three words: "Yes I am fixing it."
Update: Still Not Fixed as of May 2026
As of the time of writing — May 7, 2026 — OpenAI has not made any official announcement that the artifacting problem has been resolved. The most recent public word from the team comes from Boyuan Chen on May 4, 2026. In a reply on X, responding to a user asking whether the noise issue was still being worked on, he confirmed: "working on it" — and added that once the fix is complete, he will personally post a tweet to announce it.
So the signal to watch for is clear. Until that announcement appears, the artifact remains an open issue, the workarounds in this article remain the most practical path forward, and the fix is still in progress somewhere inside OpenAI's pipeline.
Conclusion
GPT Image 2 is, by almost every benchmark, the best image generation model OpenAI has shipped. The leap in text rendering accuracy, the expanded aspect ratios, the native reasoning, the photorealism on clean prompts — these are genuine and significant improvements over what came before. The tiling texture artifact does not erase that. But it does introduce real friction for professional workflows, particularly for anyone doing iterative generation, style-referenced work, or anything that involves re-using previously generated images as references.
The community's response has been constructive: documenting the problem systematically, developing partial workarounds, and maintaining a running collection of test cases that developers can use to track progress. And OpenAI's early responsiveness — patching the amplification bug within days, opening a direct feedback channel — suggests this is not a problem being ignored.
For now, the safest path is to treat each session as a single-image workspace, lean on fresh sessions aggressively, avoid feeding generated images back into context, and consider the API for workflows where output quality is non-negotiable. The model underneath the artifact is excellent. Hopefully, the artifact won't be its companion for long.