The image generation model index.

Reach for it on hero shots and premium fabric, where the extra detail is visible and the extra wait is affordable. It takes the same fourteen reference images as Nano Banana 2, though Google guides towards six or fewer when every reference has to be held at full fidelity. It also rendered packshots front-on when a three-quarter angle was requested, so check the framing on angled work.

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

The safe default, and the one to leave selected unless a job gives you a reason not to. Best all-round 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.

Priced for volume and it keeps the reference-image headroom of the full model, but output is 1K only. That rules it out of anything a customer orders at 2K or 4K, and leaves it a possible fast-preview engine rather than a production one.

The specialist. If the garment carries a crest, a woven label or printed lettering, this is the one engine that reads back correctly. Its output envelope is unusual: the long edge can reach 3840px, but total pixels are capped at 8.3 megapixels, so a wide 16:9 render is a true 4K while a 3:4 portrait tops out near 2450x3260. Everywhere else it is the slowest option and the most likely to refuse the job outright.

Content filtering, observed 2026-08: The strictest of the engines we run. Refuses a large share of unremarkable fashion work, including children's apparel and swimwear; a swimwear recolor returned a hard content block.

Listed for continuity. Its successor is better on both boards and is the one we ship.

A large jump over Seedream 4.5 on both boards, rendering natively at 4K and accepting up to 14 reference images, the most of anything on this page. If it inherits 4.5's tolerance for the categories other engines refuse, it replaces it outright rather than joining it.

We keep it for coverage, not for adherence. It is the weakest of the four at doing exactly what the prompt said, and it ignores the 1K tier entirely. But swimwear and lingerie are real categories with real catalogues, and this is the engine that ships them.

Content filtering, observed 2026-08: The most permissive engine we run, and the only one that completes swimwear work at 4K where the others refuse.

Listed for continuity. Cheaper than 4.5 and behind it.

Punches far above its price on both boards, and it is out on capability rather than quality. Editing takes one input image, so it cannot be handed a garment and an identity together, and native 1K output cannot serve a 4K tier.

The lowest price on this page by a wide margin. It shares the MAI line's 1K ceiling, so the same limitation applies.

The one to watch on this page. If it ships with multi-reference support it becomes a serious candidate immediately, because 2.5 already scores where it does with one hand tied.

One of the strongest editing scores anywhere on this page, and it accepts up to three reference images at up to 2K. The three-image limit is the thing to watch: it is enough for a garment plus an identity, and tight for anything more.

The previous Grok image tier. Same 2K ceiling and same three-reference limit as 2.0, at a similar price but a lower score on both boards, so 2.0 is the one worth evaluating.

On paper the strongest candidate here. It leads the editing board, renders natively at 4K, and takes up to eight reference images, which is comfortably more than our tools send in one call. Nothing structural stands in the way; it needs a bench run.

Most of 2.1's quality for a fraction of the price. If Reve earns a place, this is the value option to compare it against.

One of the best editing scores on the arena board, with no generally available API to call. Nothing to evaluate until that changes.

Historically the strongest text renderer, which is exactly the weakness GPT Image 2 covers for us. Worth re-testing the day a multi-image edit endpoint exists.

Mid-table on both boards at a mid-table price. Nothing yet that would displace what we run.

The most widely self-hosted family on this page, which matters if you run your own infrastructure. On hosted quality it sits below what we ship.

A useful reminder to read more than one board: it leads the Artificial Analysis editing ranking and does not appear on arena at all, so there is no second opinion to check it against.

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

Not comparable to anything else on this page, and deliberately so. It does not take a prompt at all: it drives a mask-based workflow that swaps the person while leaving the garment untouched. That is the opposite trade from a general engine, which regenerates everything and hopes the garment survives. Slower, and it can leave artifacts.

Built for one job: inventing a consistent model identity from a written description, with no source photograph involved. 1K only, which is all an identity reference needs.