Per generation, at a rate that depends on the model — an image made from a reference usually costs slightly more than one from a prompt alone. Each model page carries its own rate.

Build a brand kit that the next thing you make obeys.

Q: I already have a logo and colours. Can it work from those?

Yes — that is the stronger case. Hand it what you already use as the reference instead of asking it to invent one, and it becomes an enforcement mechanism for a brand you have already decided on rather than a generator of a new one.

Q: Can it get my typography right?

Partly, and this is the honest limit. A wordmark held perfectly across every asset on this page. Smaller type did not — the menu board in the café image is illegible if you look at it closely. Use generated imagery for the picture and set real text over it in whatever tool you finish in.

Q: Does the look carry into video too?

Yes — reference-driven video models take the same kind of reference and keep a subject or a palette across a clip, so the ad can be built from the same definition as the packaging.

Q: How do I stop it drifting as the set grows?

Always reference the original, never the most recent output. Editing an edit compounds whatever the last pass changed; referencing the anchor every time means the tenth asset is as close to the definition as the second one.

Q: Do I need to write code to use this?

No — Hedra Studio runs the same models on the same engine with nothing to build. Describe the brand, keep what you like, and ask for the next asset against it.

A palette in a PDF is decoration. What is worth having is a look that holds — so the ad, the packaging and the room all come back in the same two colours without being asked twice.

Start building now Build a brand kit — no code

One definition

The bag on the left was made from a written description. Everything else on this page references it and nothing else. Switch across and check the green, the circle and the wordmark.

Two honest limits: small type in generated imagery is unreliable, so treat a generated wordmark as a placeholder and set real text over it — and the colour holds its hue rather than an exact value, because each asset is lit differently.

A pink and green Goodland Coffee Roasters cup and matching folded card on a stone surface — GPT Image 2 · references the bag · nothing else

A kit is only worth having if it is enforced.

Hands holding a Goodland coffee bag at a cafe table beside a latte in a green cup — GPT IMAGE 2 · SAME REFERENCE · PALETTE IN THE SCENE

The look travels into the room, not just onto the product.

This is the same brand referenced into a café. The bag is the obvious part; look at what happened around it. The menu board, the counter front, the cup and the planter all landed in the same two colours, because the reference carried a palette rather than a logo. That is the difference between a brand kit and a sticker — one of them changes the set, the other one changes the packaging. Measured across all four assets, the green holds its hue inside about six degrees while its lightness moves with the light in the scene, which is what a real brand colour does when it is photographed rather than printed.

Three pink and green Goodland coffee bags lined up on a stone surface against a gray wall — GPT IMAGE 2 · ONE REFERENCE · FOUR ASSETS

It named the brand, and then it kept its word.

Nothing in the instruction gave this company a name. The first bag came back reading GOODLAND COFFEE ROASTERS, and because every asset after it referenced that bag, the wordmark held — same letterforms, same tracking, same position relative to the colour block — through a tumbler, a card, a café and a shelf line-up. You do not have to accept what it invents, but whatever you keep becomes the thing the rest of the set obeys.

Every model that will hold a look.

OpenAIGPT Image 2GPT Image 2.5 FlareGPT Image 2.5 Sunburst

GoogleNano Banana ProNano Banana 2Nano Banana

Black Forest LabsFlux Kontext MaxFlux Kontext ProFlux.2 [max]

ByteDanceSeedream 5.0 ProSeedream 4.5

See all models →

Built for visual inference.

Serving video is a different problem from serving text — a single request can saturate a GPU, and none of the tricks that made language models cheap apply. Hedra's engine was built for exactly that workload.

Whichever model you choose — ours or anyone else’s — it runs on the same infrastructure, behind one key. Which is what makes a look enforceable rather than aspirational: the same reference goes to an image model for the packaging and a video model for the ad, in one chain, without the asset leaving the platform in between.

Generate with API or Agent

Build with the API

One key. Keep the reference, and send it with every request that has to stay on brand.

Start building now

Define a look now.

No code — describe the brand, keep what comes back, and ask for the next asset.

Create with Agent

FAQs

What do I actually get?

Assets that share a look, and a reference that keeps them sharing it. The useful output is not a document — it is the fact that the next image, ad or clip you ask for comes back in the same palette and proportions without being re-described. Everything on this page was made that way from a single bag.

I already have a logo and colours. Can it work from those?

Can it get my typography right?

Does the look carry into video too?

How do I stop it drifting as the set grows?

How is it billed?

Do I need to write code to use this?

More you can do with Hedra

Put the look in a deckAI PITCH DECK GENERATOR →

Turn the packaging into an adAI VIDEO ADS →

hedra keys create
→ sk-hedra-••••••••

Enforce it across a catalogueDEVELOPER PLATFORM →

Decide it once.

Start building now BUILD A BRAND KIT NOW — NO CODE →