Reference · Updated monthly

The image generation model index.

What each model costs, how large it renders, how many reference images it will accept, how often it refuses ordinary work, and which ones we run in production. Ratings come from two independent leaderboards. The capability columns come from calling the endpoints ourselves.

Models from
Google4OpenAI4ByteDance4Microsoft AI4xAI2Reve2+8 more

This page compares the models themselves. If you are choosing between the platforms built on top of them, that is a different comparison: AI fashion model generators, compared.

31
Models tracked
14
Providers
11
Live in On-Model
2026-11
Last refreshed
Safe default
GPT Image 2.5

What Auto runs on most tools, and it reads back a crest or woven label correctly.

Fast drafts
Nano Banana 2

Best garment fidelity per second of wait.

Widest coverage
Seedream 5.5

Selectable on every tool, and it completes swimwear and lingerie that stricter engines decline.

Hero shots
Nano Banana Pro

Premium fabric detail when one image has to carry the page.

The short versionWhat we run today, plus what is being evaluated next. 19 more models are tracked in the full index below.
ModelUse it forEditingPrice / 1kStatusTry it
GPT Image 2OpenAILegible garment text. Superseded by 2.5.1462$211AvailableTry it →
GPT Image 2.5OpenAILegible garment text, top-tier scores.1524$211AvailableTry it →
Ideogram 4.5IdeogramA true masked editor, 2K max.1347$100AvailableTry it →
Nano Banana 2GoogleBest quality per second of wait.1387$67.0AvailableTry it →
Nano Banana 2.1GoogleGoogle's newest Flash-class model.1428$33.6AvailableTry it →
Nano Banana ProGoogleHero shots and premium fabric detail.1390$134AvailableTry it →
Seedream 4.5ByteDanceTrue 4K, loosest prompt adherence.1302$40.0AvailableTry it →
Seedream 5.0 ProByteDancePermissive and 2K-class, with tighter adherence than 4.5.1394$90.0AvailableTry it →
Seedream 5.5ByteDanceThe most realistic Seedream, on every tool.1159$20.0AvailableTry it →
FLUX 3 ImageBlack Forest LabsTen references, 4K, box-level edits.1119$48.0Candidate
Grok Imagine Image 2.0xAIStrong editing score, three references.1425$60.0Candidate
Reve 2.1ReveNative 4K, eight references.1374—Candidate

Editing score is arena.ai where available, otherwise Artificial Analysis. Jump to the full index of all 31 models ↓

Where each number comes from

Third-party leaderboards

Elo, list price and release date are pulled from the four boards below. We publish both sources side by side rather than picking a favourite, because they disagree often enough to matter.

Our own measurements

Delivered resolution and refusal behaviour for the engines we run are ours, taken by replaying identical production jobs across them. Marked ✳ in the table. Ceilings for models we have not run are the provider's own published limits.

What we will not guess

A model we have never called shows not verified rather than a repeated vendor claim. Opinion is confined to the grey line under each model and is labelled as such.

The indexRefreshed 2026-11-13

Every model worth tracking.

Grouped by provider, ranked within each group by editing score. Editing is the column that matters for production work, because every real job starts from an image you already have.

Google

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Nano Banana 2
Gemini 3.1 Flash Image
Feb 20261261 / 11431387 / 1126$67.04K14mediumAvailable

A dependable all-rounder and the pick when speed matters: the best garment fidelity per second of wait.

Content filtering, observed 2026-08: Blocks a minority of ordinary apparel work; swimwear and lingerie are the usual trigger.

Try Nano Banana 2 in On-Model →
Nano Banana 2 Lite
Gemini 3.1 Flash Lite Image
Jun 20261250 / 11101314 / 1053$33.61Kmultiplenot measuredWatching

Priced for volume, but output is 1K only, so it suits fast previews rather than production orders.

Nano Banana 2.1
gemini-nano-banana-2.1
Oct 20261328 / 11761428 / 1153$33.64K14not measuredAvailable

The successor to Nano Banana 2: up to 14 reference images, 1K to 4K output and a wider range of aspect ratios. The Google pick when you want the newest quality without Nano Banana Pro's slower render.

Try Nano Banana 2.1 in On-Model →
Nano Banana Pro
Gemini 3 Pro Image
Nov 20251248 / 11191390 / 1108$1344K14mediumAvailable

For hero shots and premium fabric, where the extra detail is worth the wait. Check the framing on angled packshots, which it tends to render front-on.

Content filtering, observed 2026-08: Behaves like Nano Banana 2; same content boundaries.

Try Nano Banana Pro in On-Model →

OpenAI

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
GPT Image 1.5
gpt-image-1.5
Dec 20251238 / 11261370 / 1118$133not verifiedmultiplenot measuredWatching

Listed for continuity. Its successors score higher on both boards.

GPT Image 2
gpt-image-2
Apr 20261383 / 11911462 / 1140$2114K ✳multiple ✳highAvailable

Reproduces crests, woven labels and printed lettering as real characters. GPT Image 2.5 does the same faster on the same envelope, so it is now the better first pick.

Content filtering, observed 2026-09: Stricter content filtering than most engines here; swimwear, lingerie and children's apparel are the usual triggers.

Try GPT Image 2 in On-Model →
GPT Image 2.5
gpt-image-2.5-sunburst
Sep 20261425 / 12291524 / 1201$2114Kmultiplenot measuredAvailable

The first engine to try when a garment carries text, or when one image has to be the best it can be. A 16:9 render is a true 4K, but total pixels stop at 8.3 megapixels, so a 3:4 portrait lands near 2480x3312.

Try GPT Image 2.5 in On-Model →
GPT Image 2.6
gpt-image-2.6
—— / 1261— / 1210$211not verifiednot verifiednot measuredWatching

OpenAI's follow-up to GPT Image 2.5, again as a high-control Sunburst model (scored here) and a faster Flare model.

ByteDance

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Seedream 4.0
seedream-4
Sep 20251140 / —1270 / ——not verifiedmultiplenot measuredWatching

Listed for continuity. Superseded by the newer Seedreams.

Seedream 4.5
seedream-4-5
Dec 20251148 / —1302 / 1039$40.04K ✳multiple ✳lowAvailable

The Seedream to pick when a swimwear or lingerie order needs a true 4K file. It follows prompts less closely than the newer Seedreams and ignores the 1K tier.

Content filtering, observed 2026-09: The two Seedreams share a content policy and are the most permissive engines we run: both complete swimwear and lingerie work that stricter engines refuse. 4.5 is the one that does it at a true 4K.

Try Seedream 4.5 in On-Model →
Seedream 5.0 Pro
seedream-5-pro
Jul 20261256 / 11031394 / 1131$90.04K ✳10 ✳lowAvailable

Outscores 4.5 and completes the same categories stricter engines decline. Total pixels stop at 4.6 megapixels and it takes ten references, so order it at 2K.

Content filtering, observed 2026-09: The same content policy as Seedream 4.5, and as permissive: it completes the swimwear and lingerie work stricter engines refuse, at 4.6 megapixels rather than a true 4K.

Try Seedream 5.0 Pro in On-Model →
Seedream 5.5
seedream-5-5
Nov 2026— / 1134— / 1159$20.04K10lowAvailable

More photographic skin and light than 5.0 Pro, garments carried faithfully and one scene held across a set. Same 4.6-megapixel ceiling and ten references, so order it at 2K.

Content filtering, observed 2026-11: The same permissive policy as Seedream 5.0 Pro: it completes swimwear work that stricter engines refuse.

Try Seedream 5.5 in On-Model →

Microsoft AI

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
MAI Image 2.5
mai-image-2.5
Jun 20261254 / 11181401 / 1124$48.11K1not measuredRuled out

Strong for its price, but editing takes one input image and output is 1K, so it cannot take a garment and an identity together or serve a 2K order.

MAI Image 2.5 Flash
mai-image-2.5-flash
Jun 2026— / 1034— / 1088$20.0not verifiednot verifiednot measuredWatching

The lightest MAI model, with the same 1K ceiling.

MAI Image 2.6
mai-image-2.6
Sep 20261333 / 11691428 / 1150$38.91.5K5not measuredRuled out

Adds editing with up to five references, which the 2.5 line lacked. Output stops near 1536x1536, short of a 2K order.

MAI Image 2.7 Preview
mai-image-2.7-preview
—n/a— / 1198—not verifiednot verifiednot measuredWatching

Among the highest editing scores on Artificial Analysis, released as a preview with no API. Nothing to evaluate until Microsoft ships one.

xAI

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Grok Imagine (Image Quality)
grok-imagine-image
Apr 2026— / 10491391 / 1098$50.02K3not measuredWatching

The previous Grok tier: same 2K and three-reference limits as 2.0, with lower scores.

Grok Imagine Image 2.0
grok-imagine-image-2
—1297 / 11681425 / 1117$60.02K3not measuredCandidate

Up to three reference images at up to 2K: enough for a garment plus an identity, tight for anything more.

Reve

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Reve 2.0
reve-2.0
Jun 20261269 / —1359 / ——not verifiedmultiplenot measuredWatching

Most of 2.1's quality at a lower tier: the value option if Reve earns a place.

Reve 2.1
reve-2.1
Jul 20261302 / —1374 / ——4K8not measuredCandidate

Renders natively at 4K and takes up to eight references in one call, enough for a garment, an identity and a pose. Nothing structural stands in the way.

Meta

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Muse Image
muse-image
—1274 / 11271403 / 1128$10.0not verifiednot verifiednot measuredWatching

A strong score on both Artificial Analysis boards. We have not evaluated it on garment work.

Ideogram

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Ideogram 4.0
ideogram-4
Jun 20261205 / 1003n/a$60.02K1not measuredRuled out

Its edit endpoint took a single image. Ideogram 4.5 added multi-image editing, so 4.5 is the one to read about.

Ideogram 4.5
ideogram-4-5
—— / 10161347 / 1065$1002K ✳5 ✳not measuredAvailable

One of the few true inpainting models here: it changes only what you describe, optionally inside a mask, and copies every other pixel. Output stops at 2048x2048, so no 4K. A natural fit for recolours and targeted changes.

Try Ideogram 4.5 in On-Model →

Alibaba

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Qwen Image 2.0 Pro
qwen-image-2
Apr 20261192 / 10331303 / —$75.0not verifiednot verifiednot measuredWatching

Mid-table on both boards. Nothing yet that would displace what we run.

Black Forest Labs

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
FLUX 3 Image
flux-3
Oct 2026— / 1115— / 1119$48.04K10not measuredCandidate

One endpoint for generation and editing: up to ten references, output up to 4K, and pixel-exact local edits placed with bounding boxes written into the prompt.

FLUX.2 [max]
flux-2
Dec 20251162 / 10211262 / 998$70.0not verifiedmultiplenot measuredWatching

The most widely self-hosted family here. On hosted quality it sits below what we ship.

Sourceful

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Riverflow 2.0
riverflow-2
Feb 2026n/an/a—not verifiednot verifiednot measuredWatching

Not currently listed on either leaderboard, so there is no score to compare it against.

Krea

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Krea 3 Ultra
krea-3
—n/a— / 1154$100not verifiednot verifiednot measuredWatching

Krea's top tier, rated on the Artificial Analysis editing board but not yet on text-to-image. We have not evaluated it on garment work.

Luma AI

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Luma UNI 1 Max
uni-1
May 20261189 / 10261334 / 1078$100not verifiednot verifiednot measuredWatching

Consistent mid-table on both boards. No fashion-specific reason to move it up the queue.

PiktID

ModelReleasedText‑to‑imagearena / AAEditingarena / AAPriceper 1k imagesMax outputRef. imagesFilteringIn On‑Model
Onda
onda
—n/an/a—4K ✳multiple ✳not measuredAvailable

Takes no prompt: it swaps the person and leaves the garment untouched, the opposite trade from a general engine. Slower, and it can leave artifacts.

Try Onda in On-Model →
Orbita
orbita
—n/an/a—1K ✳none ✳not measuredAvailable

Invents a consistent model identity from a written description. 1K only, which is all an identity reference needs.

Try Orbita in On-Model →

Two ratings per column: arena.ai first, then Artificial Analysis. They are separate populations of voters and separate scales, so compare a model against others in the same column, never one board against the other. A dash means the model is not listed on that board. ✳ marks a figure we measured ourselves.

A model is not a productWhere On-Model sits

The engine is raw material. Not the factory.

Every model on this page is a generic API call. It knows nothing about garments, colourways, SKUs, ghost mannequins, or a catalogue of four thousand images that all have to look like they were shot on the same afternoon.

What a model gives you

One image, from one prompt, with no memory of the last one. Ask it twice and you get two different lighting setups, two different models, two different interpretations of the same jacket.

What production actually needs

The same identity across 300 SKUs. A garment that survives untouched. A colourway that matches the swatch. Batch throughput, approval states, retries, and a bill that makes sense at the end of the month.

What sits in between

The prompts, the reference plumbing, the masks, the segmentation, the post-process equalisation, the QC pass. That layer is the product. The engine underneath it is swappable, which is exactly why we keep this page.

If you are comparing the platforms built on top of these engines rather than the engines themselves, that is a different question and we wrote it up separately: AI fashion model generators, compared.

In production11 engines live

What we run, and where.

Every tool exposes an engine picker. Auto is the default on all of them and resolves to the fastest capable engine, with a fallback if the chosen one refuses the job.

ToolSelectable engines
Flat-to-Model
AutoNano Banana 2Nano Banana 2.1Nano Banana ProGPT Image 2GPT Image 2.5Seedream 4.5Seedream 5.0 ProSeedream 5.5
Create Packshot
AutoNano Banana 2Nano Banana 2.1Nano Banana ProGPT Image 2GPT Image 2.5Seedream 4.5Seedream 5.0 ProSeedream 5.5
Garment Recolor
AutoNano Banana 2Nano Banana 2.1Nano Banana ProSeedream 4.5Seedream 5.0 ProSeedream 5.5GPT Image 2GPT Image 2.5
Detail Repair
AutoNano Banana 2Nano Banana 2.1Nano Banana ProSeedream 4.5Seedream 5.0 ProSeedream 5.5GPT Image 2GPT Image 2.5
Image Playground
AutoNano Banana 2Nano Banana 2.1Nano Banana ProGPT Image 2GPT Image 2.5Seedream 4.5Seedream 5.0 ProSeedream 5.5
Model Swap
AutoOndaGPT Image 2.5GPT Image 2Ideogram 4.5Nano Banana 2.1Nano Banana 2Seedream 5.5Seedream 5.0 ProSeedream 4.5
Create Identity
AutoNano Banana 2Nano Banana 2.1Nano Banana ProSeedream 4.5Seedream 5.0 ProSeedream 5.5Orbita

Switching engines is free

We bill by the size or quality tier you pick, not by engine. A render costs the same whichever model produces it, so you can pick the one that suits the job rather than the one that suits the budget.

Two engines are ours

Onda and Orbita have no leaderboard entry because they are not general-purpose models. Onda swaps the person in a photograph while leaving the garment physically untouched, which is the opposite trade from a general engine that regenerates the whole frame. Orbita invents a model identity from a written description alone.

Model by jobFor fashion and product imagery

Which model for which job.

A leaderboard tells you which image a stranger preferred. This is what each engine is actually good at once a garment is involved, and every one of them is a dropdown away inside On-Model.

GPT Image 2 for fashion imagery

Reproduces crests, woven labels and printed lettering as real characters. GPT Image 2.5 does the same faster on the same envelope, so it is now the better first pick.

Try GPT Image 2 in On-Model →6 tools · same price as every other engine

GPT Image 2.5 for fashion imagery

The first engine to try when a garment carries text, or when one image has to be the best it can be. A 16:9 render is a true 4K, but total pixels stop at 8.3 megapixels, so a 3:4 portrait lands near 2480x3312.

Try GPT Image 2.5 in On-Model →6 tools · same price as every other engine

Ideogram 4.5 for fashion imagery

One of the few true inpainting models here: it changes only what you describe, optionally inside a mask, and copies every other pixel. Output stops at 2048x2048, so no 4K. A natural fit for recolours and targeted changes.

Try Ideogram 4.5 in On-Model →1 tools · same price as every other engine

Nano Banana 2 for fashion imagery

A dependable all-rounder and the pick when speed matters: the best garment fidelity per second of wait.

Try Nano Banana 2 in On-Model →7 tools · same price as every other engine

Nano Banana 2.1 for fashion imagery

The successor to Nano Banana 2: up to 14 reference images, 1K to 4K output and a wider range of aspect ratios. The Google pick when you want the newest quality without Nano Banana Pro's slower render.

Try Nano Banana 2.1 in On-Model →7 tools · same price as every other engine

Nano Banana Pro for fashion imagery

For hero shots and premium fabric, where the extra detail is worth the wait. Check the framing on angled packshots, which it tends to render front-on.

Try Nano Banana Pro in On-Model →6 tools · same price as every other engine

Seedream 4.5 for fashion imagery

The Seedream to pick when a swimwear or lingerie order needs a true 4K file. It follows prompts less closely than the newer Seedreams and ignores the 1K tier.

Try Seedream 4.5 in On-Model →7 tools · same price as every other engine

Seedream 5.0 Pro for fashion imagery

Outscores 4.5 and completes the same categories stricter engines decline. Total pixels stop at 4.6 megapixels and it takes ten references, so order it at 2K.

Try Seedream 5.0 Pro in On-Model →7 tools · same price as every other engine

Seedream 5.5 for fashion imagery

More photographic skin and light than 5.0 Pro, garments carried faithfully and one scene held across a set. Same 4.6-megapixel ceiling and ten references, so order it at 2K.

Try Seedream 5.5 in On-Model →7 tools · same price as every other engine

Onda and Orbita are ours and work differently: Onda swaps the person in a photograph while leaving the garment physically untouched, and Orbita builds a model identity from a written description. Both are in the full index above.

What the leaderboards missMeasured September 2026

Three findings you will not find on a board.

Leaderboards measure which image a stranger prefers at a glance. They do not measure whether the logo on a jersey reads correctly, or whether the resolution you paid for is the resolution you got. We replay identical production jobs across engines to answer that.

01

Legible garment text is a family trait, not a leaderboard one

On a football kit whose crest reads FEDERAZIONE ITALIANA GIUOCO CALCIO, both GPT Image 2 and GPT Image 2.5 reproduced the lettering correctly in every sample. Nano Banana 2, Nano Banana Pro and Seedream 4.5 produced letter-shaped noise in every sample. At full-body framing everything failed, so legibility tracks the pixels the text occupies, not the tier you ordered.

02

The engines with fewer pixels won that test

The GPT Image pair got the crest right at roughly 8 megapixels while Nano Banana delivered 17 and got it wrong. More resolution is not more legibility. It is also worth checking what an engine actually returns rather than what it accepts, which is why the maximum-output column above is measured rather than quoted.

03

Content filtering is a capability, not a footnote

The strictest engine we run declines a meaningful share of unremarkable catalogue work, including children's apparel. The most permissive one is the weakest at following instructions. Swimwear and lingerie are real categories with real catalogues, so coverage and obedience trade against each other and you need both engines available.

The barWhy strong models get ruled out

A high score is not enough.

Several models near the top of both boards are marked ruled out here. That is not a quality judgement. It is a structural one, and these are the four things that decide it.

More than one reference image

Our tools send a garment, an identity and a pose reference in the same call, so a model that accepts one image cannot be asked to hold a jacket and a face at the same time, however well it scores. The ceilings vary more than you would expect: one image at the low end, fourteen at the high end, and they do not track version numbers, Seedream 5.0 Pro takes ten where Seedream 4.5 takes more.

A resolution ceiling we can state honestly

Customers buy a tier, so what matters is the gap between the tier ordered and the pixels delivered. A model capped at 1K cannot serve a 4K order at all. Others reach it only on some shapes: GPT Image 2 stops at 8.3 megapixels, so a wide 16:9 render is a true 4K while a tall portrait is not, and Seedream 5.0 Pro stops at 4.6 megapixels, so its 4K tier returns roughly 2150 pixels square. We list those ceilings rather than round them up.

A generally available API

Preview access and a leaderboard entry are not an integration. Several of the highest-scoring models on this page have no endpoint anyone outside the vendor can call.

Filtering that tolerates ordinary work

An engine that refuses swimwear, children's apparel or lingerie is not usable as a sole engine for a fashion catalogue. It can still earn a place as a specialist alongside a more permissive one.

QuestionsUpdated monthly

Asked, answered.

Which image generation model is the best?

There is no single answer, which is the reason this index exists. On the leaderboards OpenAI's GPT Image line currently sits at the top, but the newest releases move fast: GPT Image 2.6, MAI Image 2.7 Preview and Seedream 5.5 all reached the upper end of the Artificial Analysis boards within weeks of launch. On real production work the ranking changes again: the engine that renders legible lettering on a garment is not the engine that handles the widest range of subject matter, and neither is the fastest. Pick per job, not once.

How is Elo calculated, and how much should I trust it?

Both arena.ai and Artificial Analysis run blind pairwise comparisons and convert the votes into an Elo rating, the same system used for chess. It is a good measure of which output a person prefers at a glance, and a poor measure of whether a garment survived unchanged. We publish both boards side by side because they routinely disagree, and a model that only appears on one has no second opinion behind it.

Why do some models show no maximum resolution?

Because we have not called the endpoint ourselves. Capability columns are marked by how we know them: measured on our own bench, read off the provider's live input schema, or unverified. Rather than repeat a vendor claim we have not tested, we leave it blank.

Why are strong models missing from On-Model?

Almost always because they accept a single reference image. Our tools send a garment, an identity and a pose reference in one call, so an endpoint that takes one image cannot do the job at all, however well it scores. The other common reasons are a resolution ceiling below the tier a customer paid for, and no generally available API.

Which image models can edit just one part of an image?

Most engines on this page regenerate the whole frame and try to keep what you did not ask to change, so small drift elsewhere is normal. Ideogram 4.5 is one of the few built as a true masked editor: its precise-edit mode can take a mask, copies every untouched pixel exactly from your image and returns the result at your image's own size. That makes it well suited to recolours, material swaps and other targeted changes, though its output stops at 2K. FLUX 3 Image makes pixel-exact local edits too, targeted with bounding boxes written into the prompt, and renders up to 4K.

Does choosing a different engine cost more credits?

No. On-Model bills by output size, not by engine, so switching between them is free. That is deliberate: it means you can pick the engine that suits the job rather than the one that suits your budget.

How often is this page updated?

Monthly. Elo, price and release dates are pulled directly from the two leaderboards; capabilities come from our own testing and are dated where they matter. The refresh date at the top of the page is the real one, not a build timestamp.

Use all of them. Pick per job.

Every engine marked available on this page is one dropdown away inside On-Model, at the same price per image. Start with five free generations.