Can GPT Image 2 edit just one small part of a photo?

OpenAI’s GPT Image example says a user can provide a mask when they do not want the model to change a specific part of an input image. It also says that when using a mask, a prompt is still needed, meaning the mask and text instruction work together to steer the output.

The key limitation is in the same official guidance: OpenAI notes that the model might still edit some parts of the image inside the mask, even though it will avoid doing so; for an exact mask, OpenAI suggests using an image segmentation model.

So the safest read is: masks can reduce unwanted changes, but they do not guarantee that every unrequested pixel stays identical.

What OpenAI’s official material actually supports

The official sources support three narrower claims.

First, GPT Image 2 appears in OpenAI’s API model documentation. That supports saying GPT Image 2 is an OpenAI-documented model, but it does not by itself prove that every edit will preserve faces, lighting, composition, or all unmasked areas perfectly.

Second, OpenAI’s GPT Image examples include a mask workflow for editing images. The documentation says a mask can be provided if you do not want the model to change a specific part of the input image, and that a prompt is still required when using a mask.

Third, the documentation does not describe the mask as a pixel-perfect guarantee. It says the model may still edit parts of the masked area and recommends image segmentation when an exact mask is needed.

What this means for faces, lighting, and composition

If your requirement is “remove this tiny stain on a shirt, but do not change the face, skin tone, lighting, background, pose, or framing at all,” the available official evidence does not support that as a guaranteed outcome.

A mask may help the model focus the edit and avoid protected areas, but OpenAI’s wording is closer to “the model will try to avoid changing this” than “the model is mathematically prevented from changing this.”

That distinction matters for sensitive visual material: portraits, product photos, logos, packaging text, ID-style images, and brand assets. In those cases, a result that looks good at first glance may still have small changes in facial features, product edges, typography, shadows, or proportions. Because the official guidance acknowledges that masked areas can still be affected, the safer workflow is to keep the original and compare the output carefully before using it.

What developer community reports add—and what they do not prove

OpenAI Developer Community posts include several reports about gpt-image-1 mask editing problems, including poor preservation of masked areas, masks not constraining edits to a specific region, inpainting with a mask replacing the entire image, and masks appearing to be ignored.

One community reply describes GPT Image masking as prompt-based, saying the model regenerates the whole image and tries to redraw the unmasked area close to the original, rather than performing a strict pixel replacement; it also says the model may not follow the exact mask shape with complete precision.

These reports are useful as field signals, but they should be read carefully. They mainly concern gpt-image-1 and community discussion, not an official performance guarantee or failure report for GPT Image 2. Still, their direction is consistent with OpenAI’s own warning that masks may not be exact.

Be careful with “pixel-level” marketing claims

Some third-party marketing is more aggressive. For example, WaveSpeedAI describes “GPT Image 2 Edit” as a natural-language image editing model and promotes uses such as “surgical pixel-level edits” without masks, layers, or Photoshop.

That kind of language may describe a vendor’s positioning or a hoped-for user experience, but it should not override OpenAI’s own caution about masks. For the specific question “will only this tiny area change, with everything else guaranteed untouched?”, the stronger source is the official developer guidance—and that guidance does not promise pixel-level immutability.

Practical guidance before using it on important images

For rough concepts, social media variations, background experiments, and exploratory design, GPT Image mask editing can be worth testing because OpenAI does document a mask-based editing flow.

For high-precision work, use a more cautious process:

1. Write the prompt as if preservation matters. Since a prompt is still required with a mask, describe both what should change and what should remain unchanged.
2. Use better masks when boundaries matter. If the edit depends on an exact edge, OpenAI suggests using an image segmentation model to create a more precise mask.
3. Treat the output as a candidate, not a final file. The official guidance says masked areas may still be edited, so inspect faces, product contours, text, logos, and background structure before approval.
4. Compare against the original. For portraits, product listings, brand campaigns, or images with important lighting continuity, review before-and-after versions side by side rather than relying on whether the generated image looks natural.

The safest conclusion

GPT Image 2 can be considered for local image-editing workflows, but the available source-backed evidence does not support the claim that it can always change only a tiny selected area while leaving the rest of the photo completely untouched.

OpenAI’s documentation supports mask-guided GPT Image editing, and it also says masks may not be exact. For high-stakes edits, combine careful prompting, precise segmentation where needed, side-by-side comparison, and human review.