Short answer: if your main priority is preserving a person's identity through an edit, Gemini enters 2026 with a well-documented practical edge; if your priority is visually marking a region and requesting a small local change, ChatGPT offers a very direct workflow. But there is a more important caveat than picking a winner: neither should be treated as an editor with a “pixel lock” outside the selected area. OpenAI itself warns that an edit may extend beyond the selection, and independent tests available for Gemini also record collateral changes.
That changes the useful question. Instead of “which one generates the prettier image?”, the right criterion is: which one performs the requested change while keeping the face, text, objects, background and framing as close to the original as the task requires? For editorial work, product imagery, visual identity or any photo where an unrequested change can invalidate the result, fidelity matters more than creativity.
The verdict by editing task
| Priority | ChatGPT | Gemini | Practical reading |
|---|---|---|---|
| Preserve face and identity | Strong, with explicit improvements aimed at preserving appearance | Strong and particularly emphasized by Google; independent testing favors facial fidelity | Slight Gemini edge for photos of people, provided the output is checked |
| Edit a small area | Lets you select a region before making the request | Also supports circling/drawing on the target area in compatible interfaces | Functional tie; ChatGPT makes local selection especially explicit in the workflow |
| Do not change the rest of the photo | No guarantee; OpenAI says edits may extend beyond the selected area | No guarantee; tests report drift in pose, gaze, identity and alignment | Neither is safe without verification when preservation must be exact |
| Preserve existing text | Can retain or re-render it, but text is sensitive to generative reconstruction | Can also alter letters and details when reconstructing the scene | For critical text, compare with the original and run OCR or visual inspection before publishing |
| Multiple references and subject consistency | Good for conversational editing and iterative refinements | Current documentation emphasizes character/object consistency and multiple references | Gemini tends to be more attractive for multi-reference workflows |
What each company promises — and what it admits
ChatGPT: precise editing, but the selection is not a hard boundary
OpenAI launched a new generation of ChatGPT Images at the end of 2025 emphasizing “precise edits” and the ability to keep people's appearance, lighting and composition consistent between input and output. The current ChatGPT Images documentation lets you upload an existing photo, select part of it and describe the change.
The decisive detail appears in that same documentation: selection highlights are not always precise, and edits may extend beyond the selected area. That matters because a selection rectangle or brush can visually resemble a rigid Photoshop mask, while the declared behavior is not that of a protected layer. The edit remains generative.
In December 2025, OpenAI said its newer image model improved its ability to change only what was requested and preserve details such as facial likeness. That is a meaningful improvement, but it does not turn the tool into deterministic non-destructive editing. For a color correction, object removal or small clothing adjustment, the workflow is convenient; for an asset where every element outside the selection must remain identical, verification is required.
Gemini: identity consistency is an explicit priority
Google put likeness preservation at the center of the Nano Banana launch, describing the model as designed to keep people and animals recognizable while changing clothing, scenery or style. Google's official Gemini image editing page says subtle differences matter when editing someone you know and presents appearance consistency as a central goal.
Current Gemini Apps documentation says Nano Banana 2 can edit an uploaded image and can also combine several images. Google also added a way to point to the desired area by circling, drawing or annotating directly on the image. In other words, Gemini does not rely only on an abstract text description to locate the edit.
In its latest APIs, Google also documents workflows with multiple reference images and character/object consistency. That helps when the challenge is not just “change a color,” but keep the same subject across a sequence of variations.
What independent tests show in 2026
The strongest signal available is not a product claim but observed behavior when both tools are given comparable tasks.
In a photo-editing test published by DECO PROteste, ChatGPT and Gemini produced consistent results for object removal and replacement. When editing visual characteristics in people, both delivered convincing outputs; the publication observed that ChatGPT preserved more overall detail, while Gemini kept greater fidelity to the original face. That split is more useful than an absolute ranking: the better result depends on which part of the photo is “sacred.”
A Contra Labs benchmark focused on Gemini reinforces the need for caution. Across 11 production-style sessions, the model performed the requested edit in 73% of cases, kept the rest of the image stable in 64% of preservation tests, and achieved both at the same time in 55%. The sample is small and cannot support a direct comparison with ChatGPT, but it reveals the structural issue: “change one thing and preserve everything else” still fails often enough to require review.
A comparison published on August 27, 2026 by GenImagePro concluded that Gemini stands out for multiple references and subject consistency, while ChatGPT provides a particularly direct workflow for region selection and conversational refinement. It is an editorial comparison rather than a controlled scientific benchmark, so it is best treated as complementary evidence, not a definitive scorecard.
How to measure edit fidelity without fooling yourself
A useful comparison must separate aesthetic quality from fidelity. An output can look better and still be a worse edit if it changes something you did not request. A reproducible benchmark can use six checks.
- Requested change: did the requested modification happen completely and correctly?
- Protected region: how many areas outside the target show a visible change?
- Identity: do the face, hair, expression, proportions and personal features still clearly belong to the same person?
- Text: do signs, shirts, packages, screens and labels preserve exactly the same content when they were not supposed to change?
- Objects and geometry: do secondary items stay in the same place, with the same shape, count and orientation?
- Framing: do aspect ratio, crop, apparent camera position, distance and margins remain stable?
To turn this into a score, use a simple rule: mark each item as pass, fail or not applicable, and count collateral changes separately. Do not combine beauty, creativity or “overall impression” into the same score; that would hide the drift the benchmark is meant to detect.
A simple protocol to compare ChatGPT and Gemini on your own photo
Use the same original image in both tools and make one change per round. For example: “change only the red jacket to blue.” Before submitting, explicitly list what must remain unchanged: face, hair, expression, text on the cup, background, hand position, framing and lighting.
Then compare every output against the original file, not only against the version generated in the previous round. Chained edits increase the chance of accumulating small differences. If the tool offers visual selection, mark only the necessary area. If important text is present, read it literally before and after; for commercial assets, OCR or an image-difference check is worthwhile for detecting subtle changes.
A test prompt can be:
Edit only the indicated area. Change the red jacket to blue. Preserve the face, hair, expression, skin, hands, cup, any visible text, background, lighting, object positions, aspect ratio and framing without alteration. Do not add, remove, move or redraw any other element.
This kind of instruction improves the odds of preservation, but it should not be confused with a technical guarantee.
So which one should you choose?
For a photo of a person where the face is the main asset, Gemini is currently the more defensible first try. Identity preservation is an explicit product priority, and independent testing favors its facial fidelity. That still does not remove the need for inspection.
For a small local edit where you want to point to the exact area, ChatGPT is very convenient, and its selection tool reduces ambiguity in the request. The caveat is official: the change can escape the selected region, so “I marked the area” does not mean “everything else is locked.”
For critical text, product photography, rigid brand artwork or any case where a changed pixel outside the target is unacceptable, neither should be the only step. Use AI to produce the proposed edit and finish or validate it in an editor with layers, deterministic masks or image-difference comparison. In that scenario, the best tool is not the one that promises never to change anything; it is the one that creates less drift and lets you detect it before publication.
Conclusion
ChatGPT and Gemini are already capable conversational editors, but “edit without changing what I did not ask for” remains a reliability problem, not just an image-quality problem. In 2026, Gemini has an edge when facial fidelity and subject consistency are central to the task; ChatGPT stands out for the convenience of local selection and conversational refinement. For absolute preservation, however, the winning workflow is the one that includes external verification — because neither provides a perfectly rigid editing boundary.
