Nano Banana 2 vs GPT Image 2.5
GPT Image 2.5 is live in On-Model. Same job, same template, two engines: compare the light, the colour, the texture and the text on the garment.
GPT Image 2.5 is now selectable in On-Model, across Flat-to-Model, Create Packshot, Garment Recolor, Detail Repair and the Image Playground.
It is the first engine to take the top of both public preference boards off GPT Image 2. That is the headline, but it is not the reason we shipped it.
The reason is further down: it does what the template told it to do. Pick a look, and the light, the colour and the pose come back the way that look describes, instead of the engine quietly rendering its own preference and leaving you to fix it in the next round.
Three jobs it produced, from three very different briefs, all starting from flat-lay product photos:
What actually got better
Five things that were genuinely difficult before, in the order they cost us the most reshoots.
Doing what the template asked. A look that calls for hard night light used to come back as flat daylight, and the grade would wander between images in the same job. Holding to the brief is the single biggest change here, and it is the one you can see in the comparison below.
Tiny detail on the garment itself. Crests, woven labels, care tags, printed slogans, embroidery. Most engines render these as letter-shaped noise: from a distance it reads as text, and up close it is nonsense. GPT Image 2.5 holds it.
Fabric texture. Ribbing, knit structure, the difference between matte jersey and a satin finish. These survive the generation instead of being smoothed into a flat colour field.
Background colour that stays put. A set generated across several instructions used to drift, one image a half-step warmer than the next, which is exactly the kind of inconsistency that gets a batch rejected at the retailer's end.
The overall colour palette. Garment colour comes back closer to the input, and the scene grade stays where the instruction put it rather than pulling toward the engine's own house look.
And underneath all of it, it follows a template more completely. Load one of the ready-made looks, and the pose, the framing, the lens and the lighting all land closer to what the template describes, rather than the engine taking the template as a loose suggestion and rendering its own preference. Fewer instructions come back needing a second pass.
Same inputs, two engines
One job, three instructions, one identity, and the Urban Streetwear template, one of the ready-made looks that ships with On-Model. The only thing that changed between these two runs is the engine.
Only the hoodie was supplied as an input here, so ignore the rest of the styling. Watch the light, the colour and the pose.
Urban Streetwear asks for a night scene, hard directional light and attitude. Nano Banana 2 renders it as flat grey daylight in two frames out of three, and the third goes so dark the colour drains out of it. The poses are upright and static.
GPT Image 2.5 gives you the brief: amber streetlight, wet tarmac throwing the light back, real depth behind the subject, and a colour palette that stays the same warm register across all three. The poses have some intent to them, the second one especially, shot from the ground with the leg thrown toward the lens.
That is the template-adherence point in practice. Both engines produced usable pictures. Only one produced the picture the template described, three times running.
On a packshot, with no model to distract you
Same garment, same instruction, shown next to the flat-lay it came from.
Zoom into the crest and the comparison stops being subjective. The input reads FEDERAZIONE ITALIANA GIUOCO CALCIO in three tiny lines. Nano Banana 2 renders it as MEEGLATIME SUPFTHAI PLESSNVCHIBIEN: letter-shaped, correctly placed, and meaningless. GPT Image 2.5 reproduces all three lines correctly.
The colour is the second thing. The green panel comes back closer to the input, vivid rather than muted, and the navy holds its depth instead of flattening toward grey.
Texture is the third. The input is a shell fabric with a slight sheen and a fine weave. Nano Banana 2 smooths it into something closer to moulded plastic. GPT Image 2.5 keeps the weave and the sheen, and the crest keeps its raised woven edge rather than looking printed on.
On a product page those three add up to whether a shopper believes the picture.
Where to choose it
Every processing app has a Generation model control in Advanced Options, on the last step before you start the job.
Leave it on Auto and On-Model picks the best model for your set of images and instructions. Choose an engine explicitly and that is the engine that runs.
Calling it from the API
model is an optional field inside options. Everything else about the request is unchanged.
{
"project_id": "your-project-id",
"images": ["image-uuid-1", "image-uuid-2"],
"identity_code": "your-identity-code",
"instructions": [
{ "num_outputs": 2, "options": { "size": "4K", "aspect_ratio": "3:4" } }
],
"options": {
"model": "gpt_image_2_5"
}
}
Accepted values on Flat-to-Model, Create Packshot, Garment Recolor, Detail Repair and the Image Playground: "auto" (default), "nano_banana_2", "nano_banana_pro", "seedream", "seedream_5_pro", "gpt_image", "gpt_image_2_5".
Each result carries a model_used field, so a batch tells you which engine produced every image. Full reference in the Flat-to-Model API docs.
Also new: Seedream 5.0 Pro
Landing alongside it is Seedream 5.0 Pro, which matters for a different reason. It is the most permissive engine we run: it completes swimwear, lingerie and childrenswear work that stricter engines decline outright, and it scores higher than the 4.5 it succeeds.
If you have ever had a perfectly ordinary swim or intimates SKU refused mid-batch, that is the one to reach for. Seedream 4.5 is still in the picker beside it and shares the same content policy, worth knowing only if you specifically need a true 4K file, since 5.0 Pro caps lower.
In the Image Playground too
GPT Image 2.5 is available in the Image Playground, the chat-style surface for one-off work that does not fit a batch app. Generate an image from a prompt, edit an existing one, place a saved identity in a scene, then iterate turn by turn.
It is included in your On-Model subscription and billed from the same credit balance as everything else, so there is no separate seat or plan to add.
Which one, when
| If your job is | Reach for |
|---|---|
| Almost anything, on any of the processing apps | GPT Image 2.5 |
| Creating a model identity | Nano Banana Pro. Its faces read as more genuinely photographic, and GPT Image 2.5 is not offered on Create Identity |
| A large plain-fabric catalogue on a deadline | Nano Banana 2. Fastest of the lineup, and it refuses ordinary apparel work less often than the GPT Image engines |
| Swimwear, lingerie or childrenswear | Seedream 5.0 Pro |
| Anything, without wanting to decide | Auto |
Specs, preference-board ratings and measured resolution ceilings for every engine on the market are in our index of image generation models. If you want the aesthetic differences rather than the numbers, choosing an AI engine walks through them with examples.
Try it
Open any processing app, expand Advanced Options on the last step, and pick GPT Image 2.5 from the Generation model list. Same credits, same batch, same workflow.
Read Next
Choose Your AI Generation Engine
Pick from six AI engines for Flat-to-Model: Nano Banana 2 and Pro, Seedream 4.5 and 5.0 Pro, GPT Image 2 and 2.5. Or let Auto choose.
Build a Template by Describing the Shoot
Describe the shoot in plain words and On-Model plans the shots with you, then turns that plan into a reusable template you can run on every product.
Your Asset Library, Now Smarter
AI-detected keywords power the search bar so you can find any asset by content, and per-asset styling notes auto-fill the next time you reuse a garment.