FLUX.2 Pro vs Nano Banana 2 for product shots: a side-by-side

If you’ve spent any time in the image-gen leaderboards over the last month, you’ve seen FLUX.2 Pro and Nano Banana 2 trading the top spot depending on which benchmark you ran. The aggregate scores are within spitting distance. The interesting question is the one the leaderboards don’t answer: for the specific job you’re trying to do, which one wins?

We picked a job we keep running into — product photography, the kind a small brand might commission an agency for — and ran the same 30 prompts through both models.

The setup

Thirty prompts, three categories: tabletop product shots (skincare, food, small electronics), lifestyle product shots (a hand holding the object, on a desk, in a kitchen), and editing tasks (take an existing photo and swap a background, change packaging color, add a model holding the product).

Two of us rated each pair blind on a 1–5 scale across four criteria: photorealism, lighting plausibility, material accuracy (does the bottle look like glass or like CGI), and text fidelity (when the product has a label).

Where FLUX.2 Pro wins

FLUX.2 Pro is the photorealism winner outright. On tabletop shots, the soft-box reflections, the depth of field, and the material handling on glass and metal were consistently more “agency photographer” and less “3D render.”

The gap is biggest on lighting. Ask both models for “morning window light, north-facing,” and FLUX produces something a photographer would recognize. Nano Banana 2 produces something that looks lit, but lit by a neural net that has seen a lot of photographs of lit objects — slightly too even, slightly too clean.

Material accuracy is a draw on most surfaces and a FLUX win on transparent ones. Anything with a label seen through glass, anything with subsurface scattering (translucent skincare bottles, ice in a drink), FLUX is doing something subtler.

Where Nano Banana 2 wins

Nano Banana 2 is the editing winner, decisively. The 14-image multi-reference is the feature, and the feature works. Hand it a product shot, a packaging variant, and three brand references, ask for “the same shot but with the matte-finish version of the bottle on a marble countertop matching reference 3” — and on 24 of our 30 prompts the first generation was usable. FLUX.2 needed two or three iterations to converge on the same result.

Text fidelity is also a Nano Banana 2 win. When the product has a label with actual brand copy, Nano gets the kerning, the casing, and the tracking right far more often. FLUX still hallucinates a passable label, but it’s not your label.

Speed-to-output is a Nano win on every prompt — roughly 30% faster on our test runs, though that varies hour to hour with provider load.

What we actually do

For new product shots — no reference, just a prompt — we start with FLUX.2 Pro and accept that it will cost us one or two iterations to get the framing right. The hero shot at the end of the workflow is almost always FLUX.

For iteration on an existing product shot — swap backgrounds, change colors, change finishes, keep the product identical — we start with Nano Banana 2 and rarely need to switch.

So our default in the studio for product shots is: Nano Banana 2 if you’ve uploaded a reference image, FLUX.2 Pro if you haven’t. That’s roughly what the model picker recommends by default now, with one click to override.

A small caveat

This is one job category. The same comparison for, say, character illustration would probably tip more toward Nano Banana 2 (multi-ref is the killer feature for character consistency). The same comparison for photoreal scenes without a product would probably tip even further toward FLUX. The point isn’t that one model is better — it’s that asking “which model is best” without naming the job is the wrong question, and the studio exists so you can stop asking it.