Turning a piece of insight into an actual ad used to be where everything stalled — you'd know exactly what to make, then lose a week getting it made. Image generation is the bit of Atria we most enjoy showing people, because it closes that gap in about a minute. Here's how we'd work through it.

Everything on-brand comes out of your brand profile, so a thin profile means weak output. Go to Brands in the left sidebar and fill in your logo, colours, tone, USPs and audience. No brand set up yet? Raya can build one from your website URL and you tidy it up afterwards.

Then add your products — import them from a product URL or add them by hand. This one matters more than it looks. It's the difference between Raya using your product and Raya inventing a plausible-looking version of it. Products you've added to your brand are included automatically, and generated images are checked against the real product photo. If you sell several similar SKUs, give each one its own profile with its own images and description, rather than one catch-all entry.

Your logo, your call. You decide whether your logo appears in generated ads. Set the default in Brand → Visual Identity, or just tell Raya when you ask for an image.

If you already have an ad that's nearly there, go straight to number three — you're building on evidence rather than a hunch. Number one is for when the page is blank.

There's also a fifth route that runs without you: Auto-Gen, which sits under Analytics and iterates automatically on ads that are already performing. It's set when you connect an ad account, and you can switch it off from Analytics whenever you like — worth knowing it's there, because it generates in the background and draws on the same credit pool. How to turn off Auto-Gen covers it, and it's not part of the manual flow below.

Before you hit Generate, open More settings and set Variants, Aspect ratio, Scenario, Persona and Notes. We'd set aspect ratio here every single time rather than mentioning it in passing and hoping. Need something unusual? In a chat, Raya can generate any aspect ratio between 1:3 and 3:1 — just ask.

There's a model picker at the bottom-left of the generation modal, with a few options behind it, and each model comes in three quality levels: Medium, High and Max. High is the default.

You don't need to know what's under the bonnet; you do need to know they differ in quality and in what they cost you. Each option shows its per-variant cost right there in the dropdown, and hovering the Generate button gives you the total for your current model, quality and variant count — updating live as you change them. Once an image is made, its details show which quality level was used.

How we'd use that: keep the default while you're exploring directions and generating in volume, and step up to Max once you've found the concept you actually want to run. One thing to know — when you regenerate an image, it reuses the model you generated with the first time. If you want a different one, start a fresh generation. Broader detail on usage lives in Understand & manage AI credits.

Open any generated image to see it in a real phone mock-up for Feed and Reels. An optional overlay shows what the Feed crop cuts off and what the Reels interface covers, so you won't find out your headline was hidden after the ad is live.

Static ads are generated with this in mind: copy, logos, faces and products are kept inside the visible area for each Meta placement. A 9:16 version is built from the 4:5 design, so you get both sizes for the cost of one.

Custom fonts. Upload your typefaces once — .ttf, .otf, .woff, .woff2 — and your generated images stop using a generic stand-in for your brand.

The library you upload into isn't capped, but Brand fonts on the brand gives you two named roles — Headline and Body — and each one takes a font family and a weight from its own dropdown. So you're picking two faces at two specific weights, which is exactly why it pays to upload the full family: a weight that isn't in your library can't be selected. After uploading a font you can set it as your Headline or Body font straight away. You can swap what sits in those two roles for a single generation in Visual identity on the brief, and an image's details show whether your brand font actually reached the generator. One honest limitation: font uploads currently support English letters and Arabic numerals only, so non-Latin typefaces will be rejected at upload. Full walkthrough in the Custom Fonts article.

Magic Layer. When an image is almost right, this is what saves you.

Open a generated image and click the Magic Layer tool. It splits the image into its parts — text, main object, background — so you can rewrite a headline, change a colour, or move, scale or flip an element right on the canvas, with no regeneration. For anything that isn't its own layer, like words on packaging or a stray object, tap it or draw a loop around it to erase or replace just that part; those edits show their credit cost before you run them.

It works best on images with one clear subject and not much overlap. Full walkthrough in the Magic Layer article.

Our rule of thumb: small fix, use Magic Layer. Different direction, regenerate. Most of the frustration we see comes from people regenerating five times for something they could have nudged in ten seconds.

Brands built from a website URL usually get it filled in automatically, but do check it's right.

Click Save to Assets and pick a board — boards sit under a brand, so it lands where your team will look for it.

You'll get a Saved to [board] toast with a View in Assets link. Edited it in Magic Layer first? The edited version is what gets saved.

And as always — if you'd like a hand getting your brand profile into decent shape, or you've got a batch that keeps coming back wrong, just shout. Happy to look at your actual generations with you.