[Regression after July 17 update] ChatGPT image edits lose product geometry and frequently time out

Subscription

ChatGPT Pro

Summary

I am reporting a new image-generation and image-editing regression that became noticeable today, July 17, 2026. This is different from my earlier complaint about the quality difference between regular ChatGPT and Codex/Work.

Yesterday, July 16, ChatGPT was significantly better at understanding the overall form of a product and preserving that form through image-editing iterations. Today, the same type of workflow has become much less reliable. When I ask ChatGPT to edit a product image, the product itself is often changed or redrawn incorrectly, even when changing the product is not part of the request.

The editor also appears to lose or ignore its previous understanding of the product’s shape. Instead of remembering the product’s defining structure from the supplied image and the ongoing conversation, subsequent edits can treat it like a generic object and invent a different form.

At the same time, image-editing requests have become unstable. Many requests time out, remain stuck in a loading or generating state, or never finish at all.

Important: other users are reporting the same endless-loading failure

This reliability problem does not appear to be limited to my account. During the same rollout window, another ChatGPT user reported that image creation and modification repeatedly load forever and eventually end with “Streaming Interrupted”. Other users in that discussion confirmed that they were seeing the same behavior.

That report is particularly relevant because it covers both creating a new image and modifying an existing one. The symptom closely matches what I am experiencing: an image request begins, the interface continues loading without completing, and the operation eventually fails or remains unfinished.

This concurrent report does not prove a shared backend cause, but it makes the behavior less likely to be an isolated browser, conversation, or account-specific problem. Please investigate the timeout and unfinished-job behavior as a potentially broader service regression affecting image creation and editing.

Timing and possible correlation with the July 17 update

I noticed this change after the desktop update discussed by Thibault “Tibo” Sottiaux on July 17, 2026:

• Tibo’s update on X

That announcement describes changes to the ChatGPT desktop experience, including synchronized Chat/Work history and easier switching between Chat and Work modes.

I am not claiming that the announcement confirms an image-model or image-pipeline change. It does not mention one. I am reporting the timing because the image-editing behavior was noticeably better yesterday and became substantially worse after the updated experience appeared today. The timing may be related, or it may be a separate backend regression that happened during the same rollout window.

Regression 1: the product’s defining shape is no longer preserved

The most serious quality problem is loss of product identity and geometry.

When editing an existing product image, ChatGPT should understand that the product is the fixed subject. If I request a change to the background, presentation, surrounding scene, lighting, or another limited aspect, the product’s defining form should remain stable unless I explicitly ask to redesign it.

Today, the product can change unexpectedly during an edit. The output may alter its silhouette, proportions, contours, construction, orientation, structural details, or the relationship between its parts. The result may still look like a product from the same broad category, but it no longer looks like the specific product supplied in the original image.

This is especially damaging for product-design, advertising, catalog, presentation, packaging, and e-commerce workflows. A visually attractive image is not useful if the model silently changes the item being presented.

Regression 2: ChatGPT no longer retains its understanding across edits

Yesterday, ChatGPT appeared to maintain a stronger understanding of the product’s overall form during follow-up edits. It could use the existing image and conversation context as a continuing reference.

Today, that understanding appears to be lost or weakened between turns. Even within the same conversation, a follow-up edit may behave as though the product is being interpreted again from scratch. The model may preserve the general product category while forgetting the particular shape that makes this product distinct.

This makes iterative editing unreliable. Each additional request risks introducing more structural drift. Instead of making a controlled revision, the system can progressively redesign the product.

The problem is not merely that generative images vary. The task is an edit of an existing image. The supplied product image should serve as the visual source of truth. Unrequested changes to the product’s identity are a failure to preserve the editing target.

Regression 3: image-editing requests frequently time out or never finish

The image editor is also much less reliable today.

This is a major part of the regression, not a minor secondary inconvenience. The image editor cannot be used reliably when a large portion of requests never reach a final result. The concurrent user report linked above describes the same create/modify workflow becoming stuck and ending with “Streaming Interrupted,” which suggests that this failure may be affecting more than one user.

Many edit requests now exhibit one of the following behaviors:

• the request times out;
• the interface remains in a loading, editing, or generating state indefinitely;
• the progress indicator continues without producing an image;
• the edit never reaches a clear success or failure state;
• the user is left waiting without knowing whether the request is still running;
• retrying becomes necessary simply to receive any result.

This occurs while editing images in ChatGPT, not only during a complicated generation workflow. The combination of low product fidelity and frequent unfinished requests makes the feature extremely difficult to use.

Yesterday versus today

July 16, 2026

• Better understanding of the product’s overall form.
• Stronger preservation of the product across edits.
• Follow-up requests were more likely to respect the established visual subject.
• The workflow felt usable for iterative product-image editing.

July 17, 2026

• The product’s form can change during unrelated edits.
• The system appears to forget or weaken its understanding of the product between turns.
• Follow-up edits can produce a generic reinterpretation instead of preserving the supplied item.
• More requests time out or remain stuck loading indefinitely.
• The workflow is less predictable, less faithful, and less reliable than it was one day earlier.

This sudden day-to-day difference is why I believe this should be investigated as a regression rather than treated as ordinary variation between generated images.

Expected behavior

When a user edits a product image, ChatGPT should:

• treat the supplied product as the visual source of truth;
• preserve the product’s silhouette, proportions, geometry, and defining structural characteristics;
• keep unrequested parts of the product unchanged;
• remember the established product form throughout the same conversation and editing sequence;
• modify only the requested aspect whenever possible;
• avoid turning a specific product into a generic approximation of its category;
• complete the request within a reasonable time;
• return a clear and actionable error if the edit cannot be completed;
• never leave the interface loading indefinitely without a final state.

I do not expect deterministic or pixel-identical output. I do expect identity preservation and structural consistency when editing an existing product image.

Actual behavior

ChatGPT may redraw the product, alter its defining form, and lose previously established knowledge of its structure. Repeated edits can increase the amount of drift rather than preserve the original item.

In addition, a significant number of edit attempts now time out or remain in an endless loading state. Even when an edit eventually completes, the result may no longer represent the same product.

Impact as a Pro subscriber

I subscribe to ChatGPT Pro and rely on image editing as a practical creative tool. This regression creates several problems:

• product visuals cannot be trusted to remain accurate;
• every output must be checked for unintended structural redesign;
• iterative editing becomes risky because each turn can introduce further drift;
• time is wasted waiting for requests that never finish;
• retries consume additional time and potentially additional usage;
• work completed successfully yesterday may no longer be reproducible today;
• the feature is not dependable for commercial or presentation-quality product imagery.

This is more than a subjective preference about which image looks prettier. Changing the physical form of a supplied product can make the result factually wrong for the intended use.

Related reports

• Concurrent report matching the endless-loading problem: another user reported during the same rollout window that image creation or modification loads indefinitely and ends with “Streaming Interrupted.” Other users confirmed similar behavior. This is the closest public match to the reliability problem described in this issue.

• OpenAI status incident from July 7: elevated errors with image generation in ChatGPT. OpenAI marked that incident as resolved, so it does not prove that today’s behavior is the same incident. It does show that image-generation availability has recently experienced confirmed service degradation.

• A related report about edits producing a new image despite instructions to keep existing details unchanged. This is not from the same day, but it describes the same general failure mode of an edit regenerating or changing the original subject instead of preserving it.

• My earlier GitHub issue #33555 concerns the quality gap between ChatGPT and Codex/Work. This new report is different: it concerns a day-to-day regression inside the newly updated ChatGPT experience, specifically product-form preservation, conversational edit consistency, timeouts, and requests that never finish.

These reports are included as related user experiences. They do not establish that every report has the same root cause.

Requested investigation

Please investigate whether any change deployed around July 17 affected:

• the image model or model snapshot used for editing;
• routing between image-generation or editing backends;
• prompt rewriting before an image edit;
• how the existing image and conversation history are passed to the editor;
• reference-image preprocessing or compression;
• edit-locality and preservation of unrequested regions;
• session context used across multiple image edits;
• timeout, streaming, retry, or job-completion handling;
• interaction between the new Chat/Work desktop experience and image-editing jobs.

Please compare the behavior deployed on July 17 with the behavior available on July 16, particularly for product-image edits that require strong identity and geometry preservation.

Requested resolution

Please restore the previous level of product-shape understanding and edit consistency. Image edits should preserve the supplied product unless the user explicitly requests a redesign.

Please also fix the timeout and endless-loading behavior. If an image edit fails, the interface should stop within a reasonable time, explain what happened, and offer a reliable retry. It should not remain in an unfinished state indefinitely.

If the July 17 rollout intentionally changed image-generation or image-editing behavior, please disclose that change and provide a way for Pro users to select the higher-fidelity editing path.

Final note

The sudden timing is important. This workflow was meaningfully better yesterday. Today, after the updated ChatGPT desktop experience appeared, product shapes are less stable, the system appears to forget the product’s defining form during follow-up edits, and many edit requests time out or never finish.

Please treat this as a potential regression in both image quality and service reliability, not merely as normal randomness in generated images.