We ran the same trail-running shot through four image models, side by side, and expected it to settle Nano Banana vs GPT Image for people shots. We picked the cleanest frame as the winner. Our review picked a different one, because it had a point of view. And a third model had given the runner three legs, which we hadn't noticed.

The short version
  • For people and lifestyle, GPT-Image-2.6 Sunburst won our bake-off. Not because it was cleanest, but because the frame had motion and a creative direction.
  • Gemini Nano Banana 2.1 was the cleanest and was rejected as "creepy AI". We now use it for still product shots and targeted edits only.
  • FLUX 3 Image gave the runner three legs, and it slipped past our first look. Count limbs on every image from every model.
  • Higgsfield Soul 2.0 with a reference copies the reference. Wrong for a new scene, right when copying a real photo is the job.
  • Write a creative idea, not a scene inventory. A camera move, a lens, a moment, a light direction.
  • Logos start from a typeface, not a generator. And a product range is generated or shot per colourway, never recoloured from one image.
4
models, one brief, one reference image
2
three-legged runners from two different models
20/20
brand mock-ups on-brand from a typeset reference
11/13
targeted product-swap edits approved

The bake-off: one shot, four models

The brief was a photograph for a sock brand concept: a runner's lower legs on an alpine trail, thin crew socks with the brand's mark knitted into the cuff. Every model got the same written prompt, word for word, which asked for a candid 35mm film snapshot taken by a running partner, an ordinary amateur runner with dusty shins, side-on at hip height, framed from mid-thigh to feet. Every model also got the same reference image of the mark.

We ran all four through Higgsfield, the platform we use for image and video generation, and judged them side by side at full size. As of November 2026, the vendors' own names for them are:

  • GPT-Image-2.6 Sunburst from OpenAI, which its docs describe as "image generation and editing for detailed visual work".
  • Gemini Nano Banana 2.1 from Google, described as its "latest high-efficiency image generation and conversational editing model".
  • FLUX 3 Image from Black Forest Labs, which can combine up to ten reference images.
  • Higgsfield Soul 2.0, Higgsfield's in-house photo model, described on its launch page as "built for creative, fashion-aware, culture-native generation".

The results:

Model What came back Verdict
GPT-Image-2.6 Sunburst A frame with a point of view and a sense of movement Picked by our review
Gemini Nano Banana 2.1 The cleanest frame of the four, and completely static Rejected as "creepy AI"
FLUX 3 Image A runner with three legs Missed on first look, then ruled out
Higgsfield Soul 2.0 A copy of the reference image, not the scene we wrote Ruled out for this job

The review note on the first two is worth quoting in full: "Sunburst at least gave some creative direction, it looks way better than Nano Banana, that looks like a creepy AI generation because the photo has no motion or creative direction."

Then it got worse. A batch of people shots from Nano Banana 2.1, launched before the verdict came in, kept running and produced a second three-legged runner. Two models, two extra legs, one day.

We haven't used Midjourney or Figma AI, so they aren't in this comparison.

Nano Banana vs GPT Image: clean isn't the same as good

Our mistake was the scoring. We judged the frames on technical cleanliness: sharp, well exposed, no artefacts. By that measure Nano Banana 2.1 won. But a brand photograph isn't judged on whether it's free of errors. It's judged on whether anyone wants to look at it, and a frame with no motion and no point of view reads as generated in under a second, however clean it is.

GPT-Image-2.6 Sunburst's frame had the things a photographer would have chosen: where the camera sat, what was moving, what the light was doing. That's what our review was responding to.

This is the same split we described in our Claude Design walkthrough: a capable tool will produce something competent and generic unless someone gives it a direction. With image models, the model choice decides how much direction survives into the frame.

So we changed how we run a Nano Banana vs GPT Image comparison, or any other:

  • Judge point of view first. Does the frame look like somebody chose where to stand? Only then look for artefacts.
  • Judge at full size. Every frame at 640 pixels wide or larger, never a contact sheet. Thumbnails hide extra limbs entirely.
  • Count limbs before anyone else sees it. Legs, feet and hands, on every image, from every model.
  • Stop what you've ruled out. Cancel a model's queued batch the moment it loses, rather than letting it run into someone's review feed.

None of this means Nano Banana is weak. Google's own model card is candid about its limits, including that "character consistency is not always perfect between input images and generated output image", and in our hands it does a different job well. More on that below.

The four models, honestly

GPT-Image-2.6 Sunburst

Right when: people, lifestyle and anything that needs a point of view. It also produced a set of plain bottle shots for invented drinks from a written description, every label correct, approved in one round.

Wrong when: you need a mark placed exactly. On the sock concept it kept splitting the two-letter mark across the pair, one letter on the left sock and one on the right.

What people underestimate: its edit pass. Logos on clothing first came back looking stuck on. Running the image back through with a reference of the mark already knitted, and the instruction "keep everything the same, change only the logo so it looks knitted", visibly fixed it in a side-by-side test. Re-check anatomy afterwards; an edit can nudge the composition.

What people overestimate: that it follows bans. In our rounds, a ban written into the prompt never reliably stopped garment logos, scene-natural props or date stamps.

Gemini Nano Banana 2.1

Right when: still product shots, and targeted edits to an existing image. On a bridal accessories project its job was the second pass: take an on-model photograph and "change only the jewellery" to match the product photo at true size. Eleven of thirteen targeted fixes in one round were approved.

Wrong when: people from scratch. Clean, static, and the source of our second three-legged runner.

What people underestimate: how well it takes a narrow instruction on an existing image. "Keep the photo exactly; change only this" is where it earns its place.

What people overestimate: cleanliness as a proxy for quality.

FLUX 3 Image

We ran it once, in this bake-off, so this section is short.

Right when: per Black Forest Labs, jobs that combine several references or need elements placed with bounding boxes. We haven't tested either.

Wrong when: in our one test, people. Three legs.

What people underestimate: we can't say from one run.

What people overestimate: that a newer version fixes anatomy by default. Count the limbs.

Higgsfield Soul 2.0

Right when: you want a real photograph's model, framing and art direction carried over. On the bridal project, Soul 2.0 with one of the brand's own on-model photos as the reference made the base layer, which Nano Banana 2.1 then corrected for the jewellery. That two-model recipe produced the best round of the project.

Wrong when: you want a new scene built from a reference. Higgsfield's help pages explain why: "With a reference image, the prompt field becomes unavailable." Our scene prompt was never in play.

What people underestimate: that the behaviour that disqualified it in one job is the reason we chose it for another.

What people overestimate: its range. Higgsfield pitches it at fashion and editorial photography, and that's the only work we've used it for.

Which model for which job

The job Our pick Why
People and lifestyle GPT-Image-2.6 Sunburst, high quality Creative direction survives into the frame
Still product shot Gemini Nano Banana 2.1 Clean and literal, which is right for objects
Invented product from a written description GPT-Image-2.6 Sunburst Correct label lettering, consistent object across a set
Swap one element in an existing photo Gemini Nano Banana 2.1 edit Takes "change only this" well
Carry a real photo's model and framing into a new image Higgsfield Soul 2.0 with that photo as reference It copies the reference, by design
Ink drawings and watercolours Recraft V4.1 (Vector for ink, standard for watercolour) 23 of 50 candidates approved in one review
A mascot across many scenes A Nano Banana model with one master reference Held the character; needed help with pose
A logo None Set the name in type

Two rows need a footnote. The drawing row comes from a film photographer's site, where Recraft V4.1 made the ink sketches and watercolours that sit around her photographs. The mascot row is Teddy, our studio's dog, and the next section but one is about what it took to keep him on-model.

Prompting that changed the results

The bake-off also changed how we write prompts, for every model.

Write a creative idea, not a scene inventory. A list of objects gets you a frame with those objects in it. A photograph has a reason. The prompts that worked named:

  • a camera move: panning at a slow shutter, so the background streaks;
  • a lens and a height: 28 mm, low by the trail;
  • a moment: mid-air over a stream, grit kicking off the heel;
  • light with a direction: backlit sunrise, flare.

Film stock and grain helped. "Vivid", "dramatic", "editorial", "hard flash" and postcard scenery made things worse. Earlier rounds were rejected for glossy, over-muscled fitness-model legs and logos that looked "plastered on".

Say the anatomy out loud. "One person, exactly two legs", an ordinary slim amateur, natural skin with hair, dust or a scrape. Avoid crossed legs and overlapping poses; that's where extra limbs come from.

Say the product's physical facts in every prompt. The socks were one height only, a slim crew whose cuff ends at mid-calf. Leave that out of a prompt and the model drifts taller or shorter.

One reference per shot. On the Sprizzini drink drawings, a style reference was essential: a batch sent without it drew a different table. But adding a second reference ("use the table from this image") leaked that image's whole content into the frame, fruit floating in the air included. Describe props in words; give the model one image to match.

Don't let a brief accumulate. On another project a prose brief grew past 350 lines across six rounds. Each new rule pushed the model toward assembling a list instead of making a photograph. What fixed it was deleting most of the prose and passing the client's own approved photographs in as references.

Consistency across a set is the hard part

Every model is good at one image. The work is in the tenth.

Teddy, our mascot, is a 3D vinyl-toy Bernese Mountain Dog in an olive suit and terracotta tie. To get him into new scenes we use a Nano Banana model with one canonical master image as the reference and a fixed opening paragraph, pasted word for word, that describes him. Three ways that went wrong:

  • A derived reference drifts. Using a smaller copy of a past image as the reference, or rewording the opening ("stylised", "matte"), gave a generic long-snouted cartoon puppy. Always the master, never a copy.
  • A strong reference clones the pose. With the exact opening, the model copied the master's pose and sleepy sip into every scene. The review: "all the same body position and expression". Adding an explicit instruction for a completely new pose and expression, and describing both, fixed it.
  • Small defects repeat. An extra arm, a doubled glass, a tail growing from the thigh. Checked at full size, every time.

This is the limit we wrote about in the Berner Crawl build log: image models are excellent at a plausible new thing and unreliable at reproducing an exact prior thing. Consistency is a reproduction problem, so it needs a reference discipline and a review pass, not a better model.

Logos come from the typeface

The clearest result we have on generated imagery is a negative one.

For a film photographer's site, the request was a logo: a coffee glass with steam that looked like camera film, kept simple. We generated six illustrated options with an illustration model and showed four. The answer was "I hate the logos". The next round was drawn the careful way, as a pen-line cup and film strip at the stroke weight of her handwriting typeface, set beside her name. That was turned down too: "just keep the font with her name". Her name in that typeface is the logo now.

The lesson we took: when a brand already has its lettering, the logo starts from the name set in that face. Any symbol is drawn in the face's own stroke weight, one colour, at the smallest size it will be used. And any mark gets shown on its own for a yes before it's wired into a site; both rejected rounds had been integrated first, which cost far more than the drawing did.

Generators still have a job here, after the logo exists:

  • Mock-ups with real lettering. Render the wordmark in its real font with a palette strip to an image, pass it in as the reference, and ask for the letterforms "copied exactly". Twenty of twenty mock-ups came back on-brand. The lettering was close, not exact, so final artwork always comes from the font files.
  • A mark on a product. For a knitted logo, the reference that worked was a close-up of the mark already knitted into a cuff, not the clean vector. Fed the clean vector, the model tended to print the mark on rather than knit it in.
  • Test the mark on the product before building the brand around it. A mark approved on its own can still fail once it's on a sock.

Why recolouring one shot into a range fails

One shortcut looks efficient and fails on sight. For an early homepage round on the sock concept, a single generated flat-lay was recoloured eight times in code to show the range. The review rejected the whole page: "the sock displays generally are awful and look unrealistic".

The giveaway is simple. Every colourway carried exactly the same creases, folds and light, because it was the same image. No real set of socks lies that way eight times. A range is shown by generating or shooting each colourway on its own, ideally on bodies and in motion rather than flat. Plan the image count before promising a photo-led page: a page in that style needs roughly eight to twelve generated images, not one image eight times.

Our defaults

The honest answer to Nano Banana vs GPT Image is that we use both, for different jobs, and keep a third and fourth model for narrower ones.

  • People and lifestyle → GPT-Image-2.6 Sunburst, a creative idea rather than a list, limbs counted on every frame.
  • Still product shots and "change only this" edits → Gemini Nano Banana 2.1.
  • Copying a real photograph's model and framing as a base layer → Higgsfield Soul 2.0 with that photograph as the reference, then a correction pass.
  • Drawings and illustration → Recraft V4.1, with every candidate approved or denied one by one before it's used.
  • A logo → no model. The name in type, any symbol drawn at the face's stroke weight.
  • A product range → one image per colourway, never one image recoloured.

Related comparisons: AI product photography vs a real shoot: when each wins · Kling vs Seedance: what we measured animating stills

If you're choosing models for a brand's imagery, or wiring them into a workflow your team can run, that's part of our AI integration work. Tell us what you need made, and we'll tell you which parts shouldn't be generated at all.

Common questions

Is Nano Banana or GPT Image better for photos of people?

In our test, GPT Image. We gave Gemini Nano Banana 2.1 and GPT-Image-2.6 Sunburst the same trail-running brief. Nano Banana's frame was the cleanest of four models and was rejected for looking like a creepy AI image, because it had no motion or point of view. Sunburst's frame had a creative direction and won. Our rule since then is GPT-Image-2.6 Sunburst for people and lifestyle, and Nano Banana 2.1 for still product shots and targeted edits. It is one bake-off, so test on your own brief.

Which AI image model is best for product photos?

For still product shots we reach for Gemini Nano Banana 2.1, and for invented products described in words GPT-Image-2.6 Sunburst has also come back clean. The model matters less than the method: one object, a short typeset label, the same product sentence repeated word for word in every prompt. For a product you actually sell, no model should stand in for the photograph, because every model redraws the item. Use generation for scenes, variants and placeholders, and compare every frame with the real product.

Why do AI images of people look fake or creepy?

Two reasons we see repeatedly. The first is anatomy: extra legs, hands or arms, which hide at thumbnail size and show at full size. The second is a frame with no motion or point of view, which reads as generated even when it's technically clean. Prompting a creative idea instead of a list of objects helps: a camera move such as panning at a slow shutter, a lens and height, a specific moment, light with a direction, and film grain. Words like vivid, dramatic and editorial tended to make it worse.

Can AI image generators make a logo?

They can make something logo-shaped, and in our experience that's where it should stop. Generated logo options for one client were rejected outright, and a line-drawn symbol beside the name went the same way; the wordmark set in the brand's own typeface became the logo. Start from type, draw any symbol at the face's own stroke weight, and judge it at the smallest size it will be used. Generators are useful afterwards, for mock-ups, if you feed them the real lettering as a reference.

How do I keep a character consistent across AI images?

Reference the same master image every time, never a resized or cropped copy of it, and open every prompt with the same description, word for word. Using a derived still as the reference drifted our mascot into a generic cartoon face. The catch is that a strong reference also copies pose and expression, so every scene ends up looking the same. Say explicitly that you want a new pose and a new expression, describe both, and check every frame at full size for extra limbs and doubled props.

Can I recolour one product photo to show other colourways?

You can, and it will look fake. We recoloured one generated flat-lay into eight colourways in code for a sock brand concept, and the whole page was rejected for looking unrealistic. Every colour carried the exact same creases, folds and light, which no real set of products has. Generate or shoot each colourway on its own. If you sell through Google Shopping, its image rules also require the photo to reflect the item's exact colour, which a recolour can't promise.

Why does Higgsfield Soul ignore my prompt?

Because, per Higgsfield's own help pages, once a reference image is attached to Soul 2.0 the prompt field becomes unavailable. In our bake-off, Soul copied the reference and ignored the scene we had written. That makes Soul 2.0 the wrong choice for a new scene built from a reference, and a useful one when copying a real photo's model and composition is exactly what you want, such as the base layer for an on-model image you then correct in a second pass.