Every campaign needs scroll-stopping creative, and generic AI image tools fail B2B marketers in two specific ways. The first is that the output is off-brand: generic, stock-looking images that ignore your palette, your fonts, and your figure style, so they never quite fit the rest of your content. The second is worse, and it is the single most common AI brand disaster: the model hallucinates a garbled, wrong version of your company logo and drops it right into the creative.
We think image generation should be on-brand by default and constitutionally incapable of faking your logo. Here is how that works when it runs through one branded seam instead of a bolt-on.
The real problem: off-brand output and a fabricated logo
Off-brand imagery is a slow leak. Each generic image is individually fine and collectively corrosive: a feed of visuals that do not share a palette or a style reads as inconsistent, and inconsistency undercuts a brand exactly where it is trying to look established.
The fabricated logo is a fast disaster. A model asked to render "our product ad" will often try to draw the logo from memory, producing a smeared, misspelled, subtly-wrong mark. Ship that once and you have done real brand damage. The fix is not "prompt more carefully." The fix is to make it structurally impossible.
Our take: one seam, on-brand by default, logo-safe by rule
The Generate an Image Job takes a plain-language description of the scene, plus optional reference images to compose or a base image to edit. Behind it, every image the platform produces, this Job, editorial figures, cover images, in-place edits, runs through a single branded seam. That is the design decision that makes on-brand hold without per-call plumbing.
The seam layers your brand onto every generation, in order: your brand directive (voice, colors, fonts, image style), then the chosen style palette and figure style with the accent color leading, then the buyer persona. Each layer is a no-op when it is absent, so it is robust, but the default is fully on-brand.
On-brand by default, opt-out by exception. Omit the brand and the engine loads your directive automatically. The accent color leads the palette without a line of plumbing.
Never fake a logo
Beneath every scene prompt sits a guardrail, managed as an editable platform instruction rather than a hardcoded string: never fabricate or reproduce a logo, no unsolicited text, no gibberish filler. The real logo is composited into the image separately, never drawn by the model.
This is the most important rule in the system, and it is a rule, not a hope. The model is never asked to render your mark, so it cannot render it wrong. When a generated ad needs your logo, the platform places the real one.
Three ways to generate, one on-brand result
The Job runs on your chosen image model (a Gemini 3 Pro Image-class engine) with three paths, all on-brand:
- Text-to-image, from a scene description, for a fresh visual.
- Multi-image compose, when you supply reference images, to build a new image with them inside it.
- Reference-edit, when you supply a base image, to modify it while preserving most of the original.
The channel you pick sets the aspect ratio automatically: square for a feed post, landscape for a wide banner, portrait for a story.
Every image is a reusable block
The output is not just a file you download. It is filed into your content graph as its own image asset, with the prompt as alt text, editable in a full raster editor for cropping, rotation, and region edits, where a region edit itself re-runs through the same branded seam. And because it is a block, it can be reused as a reference in a future generation, composed into a carousel, or dropped into a custom graphic, always carrying the brand with it.
Describe the scene. Let the seam apply the brand. Trust the rule that the real logo is composited, never invented. That is AI imagery you can put in a campaign without a designer checking every frame for a mangled logo.
Related: put the imagery to work in a LinkedIn carousel, and animate it into a motion graphic.
