Production case study / Australia

AI fashion campaign production in Australia.

A transparent controlled test of identity consistency, simple garment fidelity, selective rerolls, and the limits that still require human review.

Controlled demonstration - not customer work

This test used a synthetic, rights-cleared, unbranded garment reference and an existing approved identity anchor. It does not report customer results, sales performance, or a biometric identity guarantee.

total candidates

Including the source garment image

campaign candidates

Generated with Nano Banana 2

first passes rejected

Rerolled against explicit QA failures

US$0.691

estimated direct cost

Across source, comparisons, and final candidates

The production question

Could one simple garment remain readable across a connected campaign?

The target was deliberately narrow: one solid-colour tailored outfit with visible construction details, one established identity, and six different campaign frames. The workflow had to preserve the cobalt colour, cropped blazer, two silver buttons, paired flap pockets, contrast-stitched lapels, and A-line midi silhouette where the camera framing allowed.

The objective was not to prove unrestricted product replacement. It was to find a credible boundary for a small paid pilot and expose where identity, styling, pose, or garment details still drift.

Workflow

Generation was one step, not the production system.

01
Define the authoritative reference
The synthetic source image, not its original text prompt, became the product specification: saturated cobalt, cropped blazer, two silver buttons, paired flap pockets, contrast stitching, and an A-line midi skirt.
02
Lock the identity and campaign direction
One approved ISLA-A11 identity anchor and the garment reference were used throughout. An accepted campaign frame was added as a third consistency reference only on selective rerolls.
03
Generate in reviewable rounds
Google Nano Banana 2 produced nine campaign candidates at 1K. The process did not treat every completed generation as usable output.
04
Reject, reroll, and document
Three first-pass candidates were rejected for skirt-length drift, identity drift, invented accessories, or weakened hair and garment consistency. Six frames passed visual review.

Selected output