GPT Image 2.5 vs Nano Banana 2 vs Muse Image: What Editing Tests Show

A model can generate a polished poster with correct words and still perform a poor local edit. Subject editing tests whether identity and protected regions survive. Text editing tests the requested characters and everything around them at the same time. No current public dataset reviewed here isolates both jobs under one controlled test, so this comparison keeps the two evidence streams separate.

Current results at a glance

The editing scores are the values displayed on the Artificial Analysis image-editing leaderboard on September 17, 2026. The arena uses blind pairwise votes on outputs made from the same source image and instruction. That makes it relevant to controlled editing, but the published table is an overall preference measure. It does not report a separate identity-preservation or text-replacement score.

The text values come from IMG.LY's model pages for Sunburst, Flare, Muse Image, and Nano Banana 2. Its pilot methodology uses identical prompts and three generations per task. A vision model transcribes the output and compares it with the required string. Those tasks include a wordmark, poster copy, product label, and CJK neon sign, but they create new images rather than edit text in an existing image.

Subject editing: the evidence favors GPT Image 2.5

A useful subject edit stays local. Changing a jacket should not reshape the face, move the hands, alter the crop, or redraw a sign in the background. Sunburst leads the blind editing arena, and Flare follows. Their displayed confidence intervals also sit above Muse Image and Nano Banana 2. That is stronger evidence than rank alone, even though the arena does not expose a subject-only slice.

OpenAI positions Sunburst for higher-precision creative tasks and Flare for quicker, general use. It also claims better reference-subject preservation and multi-turn consistency than Images 2.0. The independent arena points in the same direction for overall editing quality. It does not independently verify every product claim or prove that either variant will preserve a specific face.

Muse Image ranks above Nano Banana 2, but their intervals overlap. The current snapshot therefore supports “Muse is higher-ranked,” not “Muse will reliably beat Nano Banana 2 on your images.” Meta says Muse Image can follow natural-language or sketch instructions, combine references, and refine an output through an agentic workflow. Those are documented capabilities. The blind arena is the separate quality signal.

Nano Banana 2 remains a reasonable candidate when a job depends on several references. Google states that the Gemini app can preserve the resemblance of up to five characters and the fidelity of up to ten objects in one workflow. Capacity is not preservation quality, so that figure should not be substituted for an identity score.

Text editing: spelling is close, preservation is unmeasured

The generated-typography scores do not identify a practical winner. Flare and Muse both reached 5.0, while Sunburst and Nano Banana 2 reached 4.9. A tenth of a point in a three-run pilot is too small to choose a production workflow on its own.

The scoring method also has a production-relevant limit: it checks similarity to the requested string and does not penalize extra text. A label can score well while duplicating a word, leaving the old price in place, or inventing a second line. A reviewer still needs to check every character, punctuation mark, duplicate, omission, and unwanted phrase.

Text replacement adds collateral-damage risk. A model may spell AUTUMN STUDIO correctly while changing the typeface, sign dimensions, fingers, reflections, or background geometry. The public typography benchmark does not measure that. The editing arena may absorb some of those preferences into its overall votes, but it does not publish a text-replacement subset. Sunburst is therefore the safer starting point only as an inference from the strongest overall editing result, not as a measured text-replacement win.

What the evidence does and does not support

Run a matched subject and text editing test

Use one source photo in which a person holds a sign reading SPRING MARKET. Start every request from that same original file. Do not pass one model's output into another model.

Change only the jacket from tan to forest green.
Preserve the person's face, hair, skin tone, expression, pose, hands,
sign, sign text, background, lighting, camera angle, crop, and image dimensions.

Replace only "SPRING MARKET" with "AUTUMN STUDIO".
Keep the sign's position, size, material, type style, letter spacing,
color, shadows, and perspective unchanged. Preserve the person, hands,
clothing, background, lighting, camera angle, crop, and image dimensions.

Run each edit three times with each available variant. Record the model ID or dated snapshot, source-image hash, prompt hash, output filename, and reviewer. For the jacket edit, inspect face identity, hands and pose, framing, sign text, and every protected region. For the sign edit, require every requested character, removal of the old phrase, no duplicate text, and stable type style and surroundings.

Count an output as accepted only when every required field passes. Keep failed requests and rejected images in the record. Three attempts can expose obvious instability, but they cannot estimate a dependable pass rate. Use more source images and repetitions when a failure is expensive or the production inputs vary widely.

Download the subject and text editing test kit. It contains the full protocol and a 24-row scorecard covering two edits, four model variants, and three runs. Every row ships as NOT_RUN.

Which model should you try first?

Start with GPT Image 2.5 Sunburst for a high-stakes subject or text edit. It has the strongest current overall editing result and is the OpenAI variant positioned for tighter control. Use Flare when iteration speed matters and every output will still be reviewed. Its editing Elo is lower than Sunburst's, while its generated-text score is equally strong or stronger in the reviewed pilot.

Muse Image is the strongest alternative in this comparison. Its generated-text score matches Flare, and it ranks above Nano Banana 2 in the current editing arena. The overlapping Muse and Nano intervals mean a matched test can still reverse their order for a specific workload.

Choose Nano Banana 2 when the Gemini workflow, reference capacity, or Search grounding is a requirement that outweighs the leaderboard order. Google's Gemini image documentation covers 1K, 2K, and 4K output, multiple reference images, text rendering, and Search grounding. Those features may matter more than a small quality difference if they remove another step from the workflow.

For copy that must be legally or commercially exact, use the image model for the visual and place the final wording as editable type in a design tool. Pixel-generated lettering remains a draft until every character and every protected region has been checked.