The 11 image models, grouped by what they're actually for

The video roster differs mostly on duration. The image roster differs on things that aren't comparable to each other at all, which is why a straight table doesn't help much.

What actually differentiates them?

Four settings, and they don't all exist on every model.

References. From sixteen on the two OpenAI models down to one on Recraft 4.1. Your documentation gives sixteen as the image pool, and only two models actually offer it.

Quality. Two incompatible systems. GPT Image 2 and GPT Image 1 Mini use a quality toggle — Auto, Low, Medium, High. Everything else uses resolution tiers in K. And three models have no quality or resolution field at all.

Aspect ratios. From ten on the Nano Banana family down to four on GPT Image 1 Mini.

Batching. Only three of eleven have a variations field. The other eight generate one image.

Only three of eleven models batch. That's the setting people assume is universal and it isn't.

The batch group

Three models, and they're the only ones that generate more than one image at a time.

Recraft 4.1 is the odd one and worth dwelling on. Sixteen variations and a single reference — the most batching on the roster and the least reference capacity, which is an unusual combination.

Its reference field also reads "add an image," singular, matching the ceiling. And it's the only model in Image Create that doesn't appear in Image Edit, so anything you generate with it can't be modified with it.

Seedream 4.5's fifteen variations matches its own description — "high-volume 4K generation, fifteen in one pass." The picker and the description agree, which isn't true everywhere.

Where this falls short. A variations field means the model will produce multiple outputs in one action. It doesn't tell you whether they'll be usefully different from each other, and nothing in the picker does.

The Nano Banana group

Three models with identical references and identical aspect ratios. Only the resolution differs.

Read that carefully and Nano Banana 2 offers more resolution options than Nano Banana Pro. Same fourteen references, same ten aspect ratios, and an additional 0.5K tier at the bottom.

The 0.5K tier is the only sub-1K option anywhere on the image roster, which makes Nano Banana 2 the cheapest route to a rough draft.

Nano Banana 2 Lite shows 1K as a fixed label rather than selectable buttons, so it appears locked to that tier. Same references and aspect ratios as the other two.

This family also has the widest aspect ratio range on the image roster — ten options, including 4:5 and 5:4 that only these three offer.

The FLUX group

Three models, and reference capacity is the axis that separates them.

FLUX.2 Max and FLUX.2 Pro are identical on everything visible. Same eight references, same eight aspect ratios, no variations on either. Their descriptions differ only in intensity — "maximum-detail rendering, reference-matched" against "high-detail FLUX rendering."

The resolution field wasn't visible in the panels read, so whether that's where they separate is unknown. Rather than infer it from the descriptions, the table says so.

FLUX.2 Klein 4B has no resolution field at all, four references rather than eight, and one extra aspect ratio — it's the only FLUX model offering auto. Its description is "rapid-fire FLUX for quick exploration," which fits.

The remaining three

Each doing something none of the others does.

Three things here.

Both OpenAI models take sixteen references — the highest on the roster — and both use a quality toggle rather than resolution tiers. GPT Image 1 Mini's toggle has only Auto and Low, with no Medium or High.

GPT Image 1 Mini's description promises batching it doesn't expose. "Rapid concept batches, ten at a time," and there's no variations field in its panel. GPT Image 2 has one, at 1–10. So the batching described sits on the other model.

Seedream 5.0 Pro caps at 2K. Seedream 4.5, the older model, offers 2K and 4K. So the newer Pro model can't produce the top tier the older one can — the same inversion as the video roster's Seedance pair.

GPT Image 1 Mini's description says "ten at a time." Its panel has no variations field. GPT Image 2 has one.

What should you actually pick?

Work from the constraint.

Need to batch? Recraft 4.1 at sixteen, Seedream 4.5 at fifteen, or GPT Image 2 at ten. Those are the only three.

Need 4K? Nano Banana Pro, Nano Banana 2 or Seedream 4.5. Not Seedream 5.0 Pro, which caps at 2K.

Need a cheap draft? Nano Banana 2's 0.5K is the only sub-1K tier on the roster.

Need many references? GPT Image 2 or GPT Image 1 Mini at sixteen, then the Nano Banana family and Seedream 4.5 at fourteen.

Need 4:5 or 5:4? The Nano Banana family only.

Need text rendered in the image? GPT Image 2's description names it — "clean in-image text" — and Seedream 5.0 Pro names "native multilingual text."

All figures read from Hexcoded's Creative Studio picker in Image → Create mode in September 2026. Resolution options for FLUX.2 Max and FLUX.2 Pro were not visible in the panels read and are marked as such. Specifications change with model versions and differ by mode. Verify in the picker before relying on this.

More on model capability, access and rights in Models.