Which AI model is best for editing photos? Nano Banana, GPT Image and Muse on the same photos (October 2026)

It depends on the edit. In our eight tests, GPT Image 2.5 and Muse Image kept the original framing and pose most closely on the background swap and the relight, where both Nano Banana models reframed the portrait. On a gentle skin retouch, Nano Banana Pro and GPT Image 2.5 made subtle changes and kept both faces, Nano Banana 2.1 smoothed more and also kept them, and Muse Image sharpened the skin instead of softening it. Nano Banana Pro gave the most faithful old-photo restoration, and no model removed background people cleanly in one run. Muse Image is the cheapest at 2 credits per edit; GPT Image 2.5 costs 11 in the web app. On the independent Artificial Analysis editing leaderboard (read October 11, 2026), GPT Image 2.5 Sunburst (max) ranks first and Nano Banana 2.1 fourth; that benchmark does not measure the settings the studio uses.

We sent the same photos and the same prompts to every image and video model in AI Media Studio and kept the results unretouched. Each result shows the model, what that exact output costs, and a short note on what it did. Page reviewed October 12, 2026. Prices for every resolution and length are on the AI model price list.

The short answer: results by edit

Eight edits, one run per model and edit. A cell reads “not tested” when that run does not settle the question, for example where the prompt asked for a new scene rather than the original framing, or where we did not judge the framing of a cleanup edit. Credits are for one edit in the web app.

Image edit results by model: whether each kept the framing and the details the prompt asked to keep, and the credits it cost in the web app
Edit	Model	Kept framing and pose	Kept identity and details	Credits in the web app
Product shot: bottle on wet slate	Nano Banana 2.1	not tested	Yes: bottle and dropper recognisable	9
	Nano Banana Pro	not tested	Yes: bottle and dropper recognisable	30
	GPT Image 2.5	not tested	Yes: bottle and dropper recognisable	11
	Muse Image	not tested	Yes: bottle and dropper recognisable	2
New background: Paris street café	Nano Banana 2.1	No: pulled the camera back	not tested	9
	Nano Banana Pro	No: reframed wider	not tested	30
	GPT Image 2.5	Yes: original crop and pose	not tested	11
	Muse Image	Yes: closest to the source	not tested	2
Relight: soft window daylight	Nano Banana 2.1	Partly: head angle shifted slightly	Makeup kept	9
	Nano Banana Pro	Partly: head angle shifted slightly	Makeup kept	30
	GPT Image 2.5	Yes: pose held	Makeup kept	11
	Muse Image	Yes: closest to the source	Makeup kept	2
Add text: a label that reads GLOW SERUM	Nano Banana 2.1	Yes: rest of the photo left alone	Yes: both lines, spelled correctly	9
	Nano Banana Pro	Yes: rest of the photo left alone	Yes: both lines, spelled correctly	30
	GPT Image 2.5	Yes: rest of the photo left alone	Yes: both lines, spelled correctly	11
	Muse Image	Yes: rest of the photo left alone	Yes: both lines, spelled correctly	2
Add an object: a red ball in front of the dog	Nano Banana 2.1	Yes: framing kept	Dog kept; ball beside the paw, not in front	9
	Nano Banana Pro	Yes: framing kept	Dog kept; ball beside the paw, not in front	30
	GPT Image 2.5	Yes: framing kept	Dog kept; ball where asked	11
	Muse Image	Yes: framing kept	Dog kept; smaller ball where asked	2
Remove people behind the subject	Nano Banana 2.1	not tested	No: added a stool and crate to the street	9
	Nano Banana Pro	not tested	Partly: a pedestrian and rider were left in	30
	GPT Image 2.5	not tested	No: added cars and shop goods to the street	11
	Muse Image	not tested	Partly: redrew the shop on the left	2
Retouch skin without changing faces	Nano Banana 2.1	not tested	Faces kept (judged at reduced size)	9
	Nano Banana Pro	not tested	Faces kept (judged at reduced size)	30
	GPT Image 2.5	not tested	Faces kept (judged at reduced size)	11
	Muse Image	not tested	Partly: sharper skin; man may read older	2
Restore a damaged 1950s print	Nano Banana 2.1	not tested	Partly: sepia tint instead of black and white	9
	Nano Banana Pro	not tested	Yes: most faithful, neutral black and white	30
	GPT Image 2.5	not tested	Partly: contrast and sharpness pushed past gentle	11
	Muse Image	not tested	Yes: faithful, slightly flat	2

Which model for which job

Start from what must not change in your photo, then pick the model that held it in these runs. Each row links to the results it is based on.

Job	Try first	Why
Swap the background, keep the framing and pose	GPT Image 2.5 or Muse Image	Both kept the original crop and pose; both Nano Banana models reframed the shot. See the results
Relight a portrait without changing the pose	GPT Image 2.5 or Muse Image	Both held the pose most closely; the Nano Banana models shifted the head angle a little. See the results
Stage a product in a new scene	GPT Image 2.5 or Nano Banana Pro	All four kept the bottle. GPT Image 2.5 made the most of the light; Nano Banana Pro built the most complete room. See the results
Add a short text label	Muse Image	All four spelled both lines correctly, and Muse Image costs the least. See the results
Animate a product without altering it	Kling O3 Pro or Seedance 2.0	Both kept the bottle unchanged; Grok Imagine Video 1.5 invented a label. See the results
Animate a portrait with a visible face	Kling O3 Pro	It kept the original backdrop. Seedance 2.0 refuses real-looking faces. See the results
Follow a specific hand or camera move	Kling v3 Pro	It turned the bottle most visibly and made the strongest camera move. See the results
Walk through a room	Any of the four	All kept the furniture stable; Seedance 2.0 moved furthest into the room. See the results

Whatever the model, name what must stay in the prompt (“keep her face, makeup and jacket exactly the same”), as every prompt on this page does.

How we tested

Same source photo and the same written instruction for every model.
One run per model, first result only, no retouching or cherry-picking.
The first four image edits ran on October 9, 2026 and the last four (add an object, remove people, retouch, restore) on October 12, 2026 through the AI Media Studio API; clips ran on October 10, 2026. Each set lists its settings, and the credits shown are what each result cost.
The October 12 source photos are synthetic: fictional people and a damaged 1950s print generated with GPT Image 2.5 Sunburst, so no real person's photo was edited.
Download every run (CSV): model, prompt, settings, credits, time of run (UTC) and our verdict for each October 12 run.
“Kept framing and pose” means the crop and the subject's position match the original. “Kept identity and details” means the things the prompt said to keep (the bottle, the makeup, the label text, the faces, the street) are still there.

Image models: eight everyday edits

One run per model and edit, so treat these as examples rather than a benchmark: running the same prompt again gives a different result.

Product shot: bottle on wet slate

Prompt: Place this bottle on a wet dark slate surface with a few eucalyptus leaves, soft natural window light, premium skincare advert. Keep the bottle exactly as it is.

Settings: Edit endpoint, one uploaded photo, one output. Nano Banana 2.1 and Nano Banana Pro at 1K, GPT Image 2.5 at high quality, Muse Image at its single size. Run on October 9, 2026; results unretouched, resized to 640 px wide for this page.

Original photo: amber dropper bottle on a plain white background — OriginalThe photo every model received.

Product shot: bottle on wet slate, edited by Nano Banana 2.1: Close crop, large leaves, bottle unchanged. — Nano Banana 2.19 credits ($0.171 on Starter)Close crop, large leaves, bottle unchanged.

Product shot: bottle on wet slate, edited by Nano Banana Pro: Adds a window and bathroom setting behind the slate. — Nano Banana Pro30 credits ($0.570 on Starter)Adds a window and bathroom setting behind the slate.

Product shot: bottle on wet slate, edited by GPT Image 2.5: Strongest window light, with visible light and shadow on the wall. — GPT Image 2.511 credits ($0.209 on Starter)Strongest window light, with visible light and shadow on the wall.

Product shot: bottle on wet slate, edited by Muse Image: Wider scene with many water droplets around the bottle. — Muse Image2 credits ($0.038 on Starter)Wider scene with many water droplets around the bottle.

What the results show

All four models kept the amber bottle and dropper recognisable and followed the slate, leaves and window-light brief.
The differences are in staging: GPT Image 2.5 made the most of the window light, Nano Banana Pro built the most complete room, and Muse Image spread the most droplets for the lowest price.

New background: Paris street café

Prompt: Replace the background with a sunny Paris street café. Keep her face, makeup, hair and jacket exactly the same.

Original photo: woman with pink eye makeup and a black leather jacket against a pink and blue studio backdrop — OriginalThe photo every model received.

New background: Paris street café, edited by Nano Banana 2.1: Pulls the camera back and places her among café tables. — Nano Banana 2.19 credits ($0.171 on Starter)Pulls the camera back and places her among café tables.

New background: Paris street café, edited by Nano Banana Pro: Reframes wider and seats her at a café table. — Nano Banana Pro30 credits ($0.570 on Starter)Reframes wider and seats her at a café table.

New background: Paris street café, edited by GPT Image 2.5: Keeps the original framing; adds a café sign and the Eiffel Tower. — GPT Image 2.511 credits ($0.209 on Starter)Keeps the original framing; adds a café sign and the Eiffel Tower.

New background: Paris street café, edited by Muse Image: Keeps the original framing and pose closest to the source. — Muse Image2 credits ($0.038 on Starter)Keeps the original framing and pose closest to the source.

What the results show

GPT Image 2.5 and Muse Image kept the original crop and pose and only swapped the background.
Both Nano Banana models reframed the shot, showing more of the scene and changing her position. That reads as a new photo rather than a background swap, which matters if the framing has to stay.

Relight: soft window daylight

Prompt: Relight this portrait with soft natural daylight from a window on the left, neutral background. Keep her identity, makeup and pose unchanged.

Original photo: woman with pink eye makeup lit by pink and blue studio gels — OriginalThe photo every model received.

Relight: soft window daylight, edited by Nano Banana 2.1: Neutral daylight; head angle shifts slightly. — Nano Banana 2.19 credits ($0.171 on Starter)Neutral daylight; head angle shifts slightly.

Relight: soft window daylight, edited by Nano Banana Pro: Shows the window on the left; head angle shifts slightly. — Nano Banana Pro30 credits ($0.570 on Starter)Shows the window on the left; head angle shifts slightly.

Relight: soft window daylight, edited by GPT Image 2.5: Clear window shadows on the wall, pose held. — GPT Image 2.511 credits ($0.209 on Starter)Clear window shadows on the wall, pose held.

Relight: soft window daylight, edited by Muse Image: Soft neutral light, pose and framing closest to the source. — Muse Image2 credits ($0.038 on Starter)Soft neutral light, pose and framing closest to the source.

What the results show

Every model removed the coloured gels and produced neutral daylight while keeping the makeup.
GPT Image 2.5 and Muse Image held the original pose most closely; the Nano Banana models changed the head angle a little.

Add text: a label that reads GLOW SERUM

Prompt: Add a minimal white label to the bottle that reads "GLOW SERUM" and below it "30 ml", in a clean sans-serif font. Keep everything else unchanged.

Original photo: unlabeled amber dropper bottle on a white background — OriginalThe photo every model received.

Add text: a label that reads GLOW SERUM, edited by Nano Banana 2.1: Large label, both lines of text present. — Nano Banana 2.19 credits ($0.171 on Starter)Large label, both lines of text present.

Add text: a label that reads GLOW SERUM, edited by Nano Banana Pro: Tall label, both lines of text present. — Nano Banana Pro30 credits ($0.570 on Starter)Tall label, both lines of text present.

Add text: a label that reads GLOW SERUM, edited by GPT Image 2.5: Smaller label, both lines of text present. — GPT Image 2.511 credits ($0.209 on Starter)Smaller label, both lines of text present.

Add text: a label that reads GLOW SERUM, edited by Muse Image: Smallest label, both lines of text present. — Muse Image2 credits ($0.038 on Starter)Smallest label, both lines of text present.

What the results show

All four models placed a white label with both requested lines and left the rest of the photo alone.
The main difference is label size. At this resolution none of them misspelled the words, so the cheapest model was enough for this simple label.

Add an object: a red ball in front of the dog

Prompt: Add a red rubber ball on the rug just in front of the dog's front paws. Keep the dog, the rug and the room exactly as they are.

Settings: AI Media Studio API edit endpoint, one uploaded photo, one output. Nano Banana 2.1 and Nano Banana Pro at 1K, GPT Image 2.5 (Flare edit) at high quality, Muse Image at its single size. Run on October 12, 2026; results unretouched, resized to 640 px wide for this page. The source is synthetic: a smartphone-style photo of a golden retriever on a living room rug, generated with GPT Image 2.5 Sunburst.

Original photo: golden retriever lying on a living room rug, looking at the camera — OriginalThe photo every model received.

Add an object: a red ball in front of the dog, edited by Nano Banana 2.1: Ball beside the right foreleg rather than in front; framing and dog kept. — Nano Banana 2.19 credits ($0.171 on Starter)Ball beside the right foreleg rather than in front; framing and dog kept.

Add an object: a red ball in front of the dog, edited by Nano Banana Pro: Ball beside the right paw rather than in front; framing and dog kept. — Nano Banana Pro30 credits ($0.570 on Starter)Ball beside the right paw rather than in front; framing and dog kept.

Add an object: a red ball in front of the dog, edited by GPT Image 2.5: Ball in front of the paws; framing and dog kept. — GPT Image 2.511 credits ($0.209 on Starter)Ball in front of the paws; framing and dog kept.

Add an object: a red ball in front of the dog, edited by Muse Image: Smaller ball in front of the paws; framing and dog kept. — Muse Image2 credits ($0.038 on Starter)Smaller ball in front of the paws; framing and dog kept.

What the results show

All four added a red ball and kept the framing and the dog.
GPT Image 2.5 and Muse Image put the ball in front of the paws as asked; both Nano Banana models placed it beside a paw instead.

Remove people behind the subject

Prompt: Remove the people and the motorbike rider behind the man. Keep the man, his face, jacket and pose, and the street exactly as they are.

Settings: AI Media Studio API edit endpoint, one uploaded photo, one output. Nano Banana 2.1 and Nano Banana Pro at 1K, GPT Image 2.5 (Flare edit) at high quality, Muse Image at its single size. Run on October 12, 2026; results unretouched, resized to 640 px wide for this page. The source is synthetic: a smartphone-style half-body portrait of a fictional man on a busy street, generated with GPT Image 2.5 Sunburst.

Original photo: man in a denim jacket on a busy street with pedestrians and a motorbike rider behind him — OriginalThe photo every model received.

Remove people behind the subject, edited by Nano Banana 2.1: People and rider removed; invented a blue stool and crate. — Nano Banana 2.19 credits ($0.171 on Starter)People and rider removed; invented a blue stool and crate.

Remove people behind the subject, edited by Nano Banana Pro: Foreground people and rider removed; a mid-street pedestrian and rider remain. — Nano Banana Pro30 credits ($0.570 on Starter)Foreground people and rider removed; a mid-street pedestrian and rider remain.

Remove people behind the subject, edited by GPT Image 2.5: People and rider removed; filled in with extra cars and shop goods. — GPT Image 2.511 credits ($0.209 on Starter)People and rider removed; filled in with extra cars and shop goods.

Remove people behind the subject, edited by Muse Image: People and rider removed; shop on the left redrawn. — Muse Image2 credits ($0.038 on Starter)People and rider removed; shop on the left redrawn.

What the results show

None of the four gave a clean result: every model either changed the street or left someone in.
Nano Banana 2.1, GPT Image 2.5 and Muse Image removed the people and the rider but filled the gap with new objects or redrew a shop; Nano Banana Pro removed the people in front but left a pedestrian and a rider further back.

Retouch skin without changing faces

Prompt: Gently retouch both faces: soften under-eye shadows and even out skin tone. Do not change their faces, features, age, hair, glasses or expressions.

Settings: AI Media Studio API edit endpoint, one uploaded photo, one output. Nano Banana 2.1 and Nano Banana Pro at 1K, GPT Image 2.5 (Flare edit) at high quality, Muse Image at its single size. Run on October 12, 2026; results unretouched, resized to 640 px wide for this page. The source is synthetic: a smartphone-style portrait of a fictional couple on a city sidewalk, generated with GPT Image 2.5 Sunburst.

Original photo: smiling woman with curly dark hair and a man with glasses on a city sidewalk — OriginalThe photo every model received.

Retouch skin without changing faces, edited by Nano Banana 2.1: Smoother skin, lighter under-eyes; faces kept. — Nano Banana 2.19 credits ($0.171 on Starter)Smoother skin, lighter under-eyes; faces kept.

Retouch skin without changing faces, edited by Nano Banana Pro: More even skin, slightly reduced shadows; faces kept. — Nano Banana Pro30 credits ($0.570 on Starter)More even skin, slightly reduced shadows; faces kept.

Retouch skin without changing faces, edited by GPT Image 2.5: Subtle retouch; faces unchanged. — GPT Image 2.511 credits ($0.209 on Starter)Subtle retouch; faces unchanged.

Retouch skin without changing faces, edited by Muse Image: No softening: sharper skin and more visible lines; man may read slightly older. — Muse Image2 credits ($0.038 on Starter)No softening: sharper skin and more visible lines; man may read slightly older.

What the results show

Nano Banana 2.1, Nano Banana Pro and GPT Image 2.5 retouched the skin and kept both faces. Nano Banana 2.1 gave the smoothest skin and lightest under-eyes; Nano Banana Pro and GPT Image 2.5 made subtler changes.
Muse Image did the opposite of the request: it sharpened the skin and made lines more visible.
We judged the faces at reduced size, not pixel by pixel, so check small details on your own photo.

Restore a damaged 1950s print

Prompt: Restore this old photograph: remove the scratches, the crease, the dust spots and repair the torn corner. Improve contrast gently. Keep it black and white and keep both people's faces, clothes and poses exactly as they are.

Settings: AI Media Studio API edit endpoint, one uploaded photo, one output. Nano Banana 2.1 and Nano Banana Pro at 1K, GPT Image 2.5 (Flare edit) at high quality, Muse Image at its single size. Run on October 12, 2026; results unretouched, resized to 640 px wide for this page. The source is synthetic: a scan of a damaged 1950s black-and-white print of a fictional couple, generated with GPT Image 2.5 Sunburst.

Original photo: faded black-and-white print of a couple in front of a brick house, with scratches, a diagonal crease and a torn corner — OriginalThe photo every model received.

Restore a damaged 1950s print, edited by Nano Banana 2.1: Damage removed; a warm sepia tint remains. — Nano Banana 2.19 credits ($0.171 on Starter)Damage removed; a warm sepia tint remains.

Restore a damaged 1950s print, edited by Nano Banana Pro: Damage removed; most faithful, neutral black and white. — Nano Banana Pro30 credits ($0.570 on Starter)Damage removed; most faithful, neutral black and white.

Restore a damaged 1950s print, edited by GPT Image 2.5: Damage removed; contrast and sharpness boosted, looks less like a period print. — GPT Image 2.511 credits ($0.209 on Starter)Damage removed; contrast and sharpness boosted, looks less like a period print.

Restore a damaged 1950s print, edited by Muse Image: Damage removed; faithful, slightly flat. — Muse Image2 credits ($0.038 on Starter)Damage removed; faithful, slightly flat.

What the results show

All four removed the scratches, crease, dust and torn corner.
Nano Banana Pro gave the most faithful result in neutral black and white, and Muse Image was also faithful, if a little flat. Nano Banana 2.1 left a sepia tint, and GPT Image 2.5 pushed contrast and sharpness so the photo looks less like a period print.

Video models: four photos brought to life

Clips play muted and loop. Veo 3.1 and Sora 2 are no longer offered, so they are not shown.

Product orbit

Prompt: Slow camera orbit around the bottle on the table, soft studio light glinting off the glass, premium skincare advert. Keep the bottle and its label exactly as they are.

Settings: Image-to-video from the photo shown, 5 seconds, audio off. Kling O3 Pro: no resolution option; Grok Imagine Video 1.5: 720p; Kling v3 Pro: no resolution option; Seedance 2.0: 720p. Run on October 10, 2026; clips unretouched, resized to 540 px for this page and shown without sound.

Source still: amber dropper bottle on a plain white background — OriginalThe photo every model received.

Kling O3 Pro85 credits ($1.61 on Starter)Subtle movement; the bottle stays as it is.

Grok Imagine Video 1.5105 credits ($1.99 on Starter)Turns the bottle and reveals a white label that is not in the photo.

Kling v3 Pro85 credits ($1.61 on Starter)The strongest camera move: a push-in toward the bottle.

Seedance 2.0220 credits ($4.18 on Starter)Subtle movement; the bottle stays as it is.

What the results show

Kling O3 Pro and Seedance 2.0 kept the bottle unchanged, as the prompt asked.
Grok Imagine Video 1.5 invented a label on the back of the bottle, so check product details when you use it for ads.
None of the four produced a full orbit; Kling v3 Pro moved the camera the most.

Portrait comes alive

Prompt: She turns toward the camera, smiles and tucks her hair behind her ear, a gentle breeze, natural daylight. Keep her face, makeup and jacket unchanged.

Source still: woman with pink eye makeup and a black leather jacket against a pink and blue backdrop — OriginalThe photo every model received.

Kling O3 Pro85 credits ($1.61 on Starter)Smiles and touches her hair; the pink and blue backdrop stays.

Grok Imagine Video 1.5105 credits ($1.99 on Starter)Smiles and plays with her hair; the backdrop turns beige.

Kling v3 Pro85 credits ($1.61 on Starter)Smiles and touches her hair; the backdrop turns light grey by the end.

Declined by the model

Seedance 2.0No charge (declined)Declined: Seedance 2.0 refuses photos that show a real-looking face, whatever the prompt.

What the results show

All three clips show the smile and hair gesture from the prompt and keep the makeup and jacket.
Kling O3 Pro kept the original backdrop; Grok Imagine Video 1.5 and Kling v3 Pro replaced it with lighter tones.
Seedance 2.0 refused this photo (likenesses of real people), and refused jobs cost no credits. Photos of people without a visible face work, as the Hands set shows.

Hands with product

Prompt: She gently turns the bottle in her hands to show it from the side, soft natural light, premium skincare advert. Keep the bottle exactly as it is.

Source still: two hands holding an amber dropper bottle over a marble bathroom counter, no face visible — OriginalThe photo every model received.

Kling O3 Pro85 credits ($1.61 on Starter)Small, steady hand movement.

Grok Imagine Video 1.5105 credits ($1.99 on Starter)Small, steady hand movement.

Kling v3 Pro85 credits ($1.61 on Starter)Tilts the bottle sideways, the clearest take on the prompt.

Seedance 2.0220 credits ($4.18 on Starter)Camera moves in closer as the hands lift the bottle.

What the results show

All four models kept the bottle and the bathroom setting consistent through the clip.
Kling v3 Pro followed the turn-the-bottle instruction most visibly; Seedance 2.0 added a push-in.

Room walkthrough

Prompt: Smooth slow dolly forward into the living room, warm afternoon sunlight moving across the sofa and floor. Keep the furniture and layout unchanged.

Source still: warm Scandinavian living room with a beige sofa, round mirror and wooden coffee table — OriginalThe photo every model received.

Kling O3 Pro85 credits ($1.61 on Starter)Gentle dolly; furniture unchanged.

Grok Imagine Video 1.5105 credits ($1.99 on Starter)Gentle dolly; furniture unchanged.

Kling v3 Pro85 credits ($1.61 on Starter)Gentle dolly; furniture unchanged.

Seedance 2.0220 credits ($4.18 on Starter)The deepest move into the room; furniture unchanged.

What the results show

Every model kept the furniture and layout stable, so all four suit simple room walkthroughs.
Seedance 2.0 travelled furthest into the room; the other three made a gentler move.

What each edit costs in the web app

Dollar amounts are the plan's price per credit: $19 for 1,000 credits a month on Starter and $199 for 20,000 on Ultra. The web app runs GPT Image 2.5 at high quality; low and medium quality are available through the API. Higher resolutions and every video length are on the AI model price list.

Credits and dollar cost of one image edit per model and setting
Model	Setting	Where	Credits per edit	On Starter	On Ultra
Muse Image	one size (about 2.5 MP)	Web app	2	$0.038	$0.020
Nano Banana 2.1	1K	Web app	9	$0.171	$0.090
Nano Banana Pro	1K or 2K	Web app	30	$0.570	$0.298
GPT Image 2.5	high quality	Web app	11	$0.209	$0.109
GPT Image 2.5	low quality	API only	2	$0.038	$0.020
GPT Image 2.5	medium quality	API only	3	$0.057	$0.030

What the benchmarks say

Artificial Analysis's AA-Image-Editing v2.0 leaderboard ranks editing models by Elo from blind votes from a recruited human panel plus public Image Arena votes cast before January 1, 2026. We read it on October 11, 2026; scores for the models on this page:

AA-Image-Editing v2.0 Elo scores for the models in this comparison, read October 11, 2026
Model variant on the leaderboard	Elo (95% interval)	Rank
GPT Image 2.5 Sunburst (max)	1184 ±7	1
GPT Image 2.5 Flare (max)	1165 ±7	2
Nano Banana 2.1	1140 ±9	4 (range 3–4)
Muse Image	1121 ±7	7 (range 6–8)
Nano Banana Pro (Gemini 3 Pro Image)	1103 ±7	14 (range 13–14)

The leaderboard judges preference on its own photos and instructions, for the exact variants named above, such as GPT Image 2.5 at max quality. It does not measure the settings the studio uses (GPT Image 2.5 at high quality, Nano Banana at 1K), and it does not check whether an edit kept the original framing, which is where the models differed most in our runs. Source: Artificial Analysis, AA-Image-Editing v2.0.

Run these models on your own photo

Open the editor with a model already picked, upload a photo and write what to change and what to keep. The credit cost is shown before you generate. Every edit needs a paid plan. Plans start at $19/month.

New to prompt editing? Start with the AI image editor guide, or see the model guides for what each model is best at.

Frequently asked questions

Which AI image model is best for editing a photo?

It depends on the edit. In these tests GPT Image 2.5 and Muse Image kept the original framing and pose most closely on the background swap and the relight, while both Nano Banana models reframed the portrait. GPT Image 2.5 and Muse Image also put an added object exactly where we asked. For a gentle skin retouch, Nano Banana 2.1, Nano Banana Pro and GPT Image 2.5 kept the faces and Muse Image did not soften the skin; for restoring an old print, Nano Banana Pro was the most faithful. For a simple text label all four did the job, and Muse Image costs the least at 2 credits per edit.

Which AI model keeps a face unchanged when retouching?

In our skin retouch (soften under-eye shadows, even out skin tone, change nothing else), Nano Banana Pro and GPT Image 2.5 made subtle changes and kept both faces. Nano Banana 2.1 smoothed the skin more and also kept the faces. Muse Image sharpened the skin and made lines more visible instead of softening them, and the man may read slightly older. We judged the faces at reduced size, on one synthetic photo, so check fine details on your own photo before you rely on it.

Which AI model is best for restoring an old photo?

In our test on a synthetic damaged 1950s print, all four models removed the scratches, crease, dust and torn corner. Nano Banana Pro was the most faithful and kept a neutral black and white; Muse Image was also faithful but slightly flat. Nano Banana 2.1 left a warm sepia tint, and GPT Image 2.5 boosted contrast and sharpness so the result looks less like a period print. Nano Banana Pro costs 30 credits per edit at 1K and Muse Image 2.

Can AI remove people from the background of a photo?

All four models removed most of the people behind our subject, but none gave a clean result in one run. Nano Banana 2.1 invented a stool and crate, GPT Image 2.5 filled the gap with extra cars and shop goods, Muse Image redrew a shop, and Nano Banana Pro left a pedestrian and a rider further back. Expect to check the background and run the edit again.

Which AI video model keeps a product unchanged?

On the product orbit, Kling O3 Pro and Seedance 2.0 kept the bottle as it was. Grok Imagine Video 1.5 added a label that is not in the photo, so check product details before using a clip in an ad.

Why is there no Seedance 2.0 clip for the portrait?

Seedance 2.0 refuses photos that show a real-looking face, whatever the prompt, and a refused job costs no credits. It animated the photo of hands holding a bottle, where no face is visible.

Is Nano Banana better than GPT Image for photo editing?

Not across these eight edits. GPT Image 2.5 kept the original crop on the background swap, held the pose on the relight and put the added ball where we asked, where Nano Banana 2.1 pulled the camera back, shifted the head angle slightly and placed the ball beside a paw. On the skin retouch both kept the faces (Nano Banana 2.1 smoothed more), and on the restoration Nano Banana 2.1 left a sepia tint while GPT Image 2.5 pushed contrast. Nano Banana 2.1 costs 9 credits per edit at 1K and GPT Image 2.5 costs 11 in the web app. AA-Image-Editing v2.0 also ranks GPT Image 2.5 higher (1184 for the Sunburst max variant against 1140 for Nano Banana 2.1, read October 11, 2026).

Which model is best for adding text to an image?

For a short label, any of the four. All of them added a white label reading "GLOW SERUM" and "30 ml" with both lines spelled correctly and left the rest of the photo alone; the main difference was label size. Muse Image did it for 2 credits. We have not tested longer text or posters yet.

What is the cheapest AI model for photo edits?

Muse Image, at 2 credits per edit in the web app ($0.038 on Starter). GPT Image 2.5 at low quality also costs 2 credits, but low and medium quality are API options; the web app runs GPT Image 2.5 at high quality for 11 credits.

Is Nano Banana Pro worth the extra cost over Nano Banana 2.1?

Only for some edits. Nano Banana Pro costs 30 credits per edit at 1K or 2K against 9 for Nano Banana 2.1 at 1K. It gave the most faithful old-photo restoration, where Nano Banana 2.1 left a sepia tint, and it built the most complete room on the product shot. But it reframed the background swap, shifted the head angle on the relight and misplaced the added ball just like Nano Banana 2.1, and both kept the faces on the skin retouch.

What model does Canva Magic Edit use?

Canva says Magic Edit is powered by Google's Nano Banana model (canva.com/features/ai-photo-editing, checked October 11, 2026). Canva does not name the version on that page, so we cannot say which Nano Banana model it runs.

Can I use these models without separate subscriptions?

Yes. Every model on this page runs in AI Media Studio on one plan that uses credits, so you do not need a separate account with each model maker. Plans start at $19 a month for 1,000 credits, and the cost of each run is shown before you generate.

Can I use the edited photos commercially?

Paid AI Media Studio plans include a commercial license. You need the rights to the photos you upload.

Dollar amounts in captions use the Starter plan ($19 for 1,000 credits a month); the price list also shows Ultra.