Can GPT Image 2 edit only part of a photo?

Quick take

What the official docs actually prove

OpenAI’s API Reference includes a Create image edit method, and OpenAI’s Image generation guide describes Edits as a way to modify existing images. OpenAI’s API documentation also includes a GPT Image 2 model page.

That is enough to support a conservative claim: GPT Image 2 belongs in OpenAI’s image-generation and image-editing ecosystem. But “there is an edit workflow” is not the same as “the model will preserve all non-selected pixels exactly.” The cited official documentation does not, by itself, justify that stronger promise.

Why a mask is guidance, not a hard selection

A mask can tell a generative model where to focus. It should not be treated like a traditional photo-editing selection that the model cannot cross.

The OpenAI Developer Community discussion on GPT Image 2 masking says the gpt-image-2 API supports the mask field. That matters: masks are relevant to the workflow. But other developer reports are more cautious. One report says an images.edit mask did not constrain the edit to the intended area, and another response says GPT Image masking is prompt-based guidance and may not be followed exactly.

The strongest caution comes from evaluation evidence. An arXiv graphic-design benchmark reports that GPT-Image models failed to reliably confine masked edits to the masked region. That does not mean every localized edit will fail. It means the safe expectation is: masks can improve control, but they are not a pixel-level guarantee.

Where this matters in real work

1. Replacing a background

Background swaps are a reasonable use case for an image-editing workflow. But after generating the new image, do not check only whether the background changed. Compare the output with the original and inspect the product edge, silhouette, shadows, reflections, color, scale and whether the subject has been subtly reinterpreted.

Because masked edits may affect areas outside the mask, side-by-side comparison is essential.

2. Changing packaging or labels

GPT Image 2 can be useful for packaging mockups, creative directions and quick variants. The risk is not just that the label change may be imperfect. The model may also alter a logo, printed text, proportions, texture, nearby objects or other details you expected to preserve.

That is why a mask should not be described as a guarantee of exact preservation for packaging assets.

3. Fixing one small area

For blemishes, small object changes or local cleanup, use a clear mask and a precise prompt. A safer prompt describes both what should change and what must stay the same—for example, the person’s facial features, the product shape, the brand mark, printed text, background and lighting consistency.

Even then, treat the result as generative editing. Check the unmasked regions after the output is created.

A safer workflow to reduce unwanted changes

1. Keep the original master file. Do not overwrite the source image. You need an untouched reference for rollback and comparison.
2. Change one thing at a time. Swap the background first, then adjust packaging, then fix small details. Stacking too many instructions makes drift harder to diagnose.
3. State what must be preserved. Do not only say what to edit. Also specify the product, logo, text, proportions, colors, shadows and background elements that should remain unchanged.
4. Use the mask, but do not rely on it alone. Public evidence does not support treating the mask as a hard pixel boundary, so pair it with explicit preservation instructions.
5. Compare output against the original. Look closely at unselected areas, faces, product geometry, typography, brand elements and packaging details.
6. Do manual QA before final delivery. If the asset needs exact branding, legally important packaging text or pixel-level consistency, a designer or retoucher should review and, if necessary, finish the file manually.

The safest way to describe the capability

A defensible external wording is:

GPT Image 2 supports image-editing workflows and can be used with mask-guided prompts for localized changes, but public evidence does not support guaranteeing that only the selected pixels, background, label or object will change while everything else remains identical.

Avoid saying that GPT Image 2 can guarantee a background, package or selected spot will change while the rest of the image stays perfectly untouched. For commercial product images, treat it as a fast generative editing tool—not as a final, pixel-preserving retouching step—and always QA the areas you did not intend to edit.