Can AI keep a brand consistent across every asset?

Yes. In FLORA you make one approved reference image, then generate every other asset from it with an image-to-image model, so the product, label, palette and light carry over. In our run, three new formats made from one reference kept the same tin and label; the same request without the reference came back as a green tin with invented copy.

At a glance

Key features

• Reference images: Every new asset is generated from the approved image, not from words alone.
• Batch formats: Up to 20 formats or scenes go to FLORA in one call and finish together.
• Canvas lineage: Each asset sits on the canvas wired to the reference it came from.
• Audit skill: flora-brand-consistency-audit judges a set against written rules and names the image that broke one.

Best for

• Turning one approved hero shot into banners, stories and social posts
• Small brands without a studio that still need a set that hangs together
• Seasonal refreshes where the product must not change
• Agencies producing many formats for one client at once

Step by step

1. Step 1: Approve one reference image

   Generate or upload the image that defines the brand: the product as it really looks, the palette, the light. Everything after is judged against it, so check the label, color and proportions before you move on.

2. Step 2: Write the rules down

   Name what must hold in every asset: product identity, palette, lighting direction, allowed surfaces. Put those words in every prompt and keep them for the audit later.

3. Step 3: Generate every format from the reference

   Ask for all the formats in one message. The agent sends each one to an image-to-image model with the reference attached.

   Using the reference image as the brand standard, make three new assets with FLORA, GPT Image 2.5 Flare:
   1. 16:9 web banner, tin on the right third, empty linen-toned space on the left.
   2. 9:16 story, the tin held in two hands over a wool blanket.
   3. 1:1 flat lay, lid off, loose tea visible, brass scoop beside it.
   Keep the same tin, label, single soft window light from camera left and warm palette of terracotta, cream and oat.

4. Step 4: Audit the set before anyone sees it

   Run the flora-brand-consistency-audit skill on the finished set with your written rules. Rerun only the images it flags.

What happened when we held one brand across assets in FLORA?

On October 16, 2026, we made a reference plate for a made-up tea brand with t2i-gpt-image-2-5-flare at 1:1: a round terracotta tin with a cream label reading HEARTH TEA, a ceramic cup beside it, undyed linen, soft window light from the left. It took 34 seconds and cost $0.055. We had no real brand to use, so the plate stands in for a client's approved product shot.

The agent then sent three generations in one flora_create_generations call with model is2i-gpt-image-2-5-flare, the plate in params.image_urls, and aspect ratios of 16:9, 9:16 and 1:1. Each took 35 to 39 seconds and cost $0.064. The banner put the tin on an oak counter beside a steaming kettle, the story put it in two hands over a knit blanket, and the flat lay showed it open with a brass scoop. Tin shape, color and label matched the plate in all three.

As a control, we sent the flat lay request with no reference to t2i-gpt-image-2-5-flare ($0.055). It returned a dark forest green tin with a green lid and a label reading EARL GREY and BOTANICALS · HERITAGE · WELLNESS. It kept the brand name and changed the product.

What went wrong: the reference plate had small label lines we never asked for, LOOSE LEAF TEA, SMALL BATCH • NATURAL INGREDIENTS and a net weight. Every reference-based image copied them faithfully. The fix is to put the exact label copy in the reference prompt, or correct the plate, before generating the set.

Models

t2i-gpt-image-2-5-flare, is2i-gpt-image-2-5-flare

Cost

$0.055 plate, $0.064 per reference-based image, $0.247 for the set

Time

35 to 39 s per image, parallel

Images

1 reference, 3 formats, 1 control without reference

Reference image or prompt only: pros and cons

Why does a reference image beat a longer prompt?

Words leave room. "A terracotta tea tin with a cream label" describes thousands of tins, and a model picks a new one every time. A reference image fixes the exact tin, so the prompt only has to describe what changes: the scene, the crop, the hands holding it.

What should the written rules cover?

Write rules that two people would apply the same way. They go into every prompt and into the audit afterwards.

• Product identity: shape, color, label wording and mark.
• Palette: the colors allowed to dominate, and the ones that are not.
• Lighting: direction, softness, no flash.
• Surfaces and settings that are allowed.
• Output spec: aspect ratios and sizes each channel needs.

How do I check a whole set at once?

Ask your agent to run the flora-brand-consistency-audit skill with your rules. It describes each image, transcribes its text and judges every rule with evidence, at about $0.01 per image in our run. The audit guide covers it step by step.

Frequently asked questions