Keeping AI images on brand: reference images that actually work

The first AI image you generate for your brand is usually great. The tenth is the problem. Same prompt style, same model, and somehow the product is a slightly different shape, the mascot's ears changed, the palette wandered from warm to cool. Nothing is wrong exactly. It just does not look like it came from the same company.

Consistency is a workflow property, not a prompting trick. These are the pieces that actually hold a look together.

Reference images beat descriptions

You cannot describe your product accurately enough in words. A reference image is worth the proverbial thousand: most current image models accept one or more reference images alongside the prompt, and they treat them as ground truth for what the subject looks like. In the image studio, dock your product shot or character art as a reference, then let the prompt describe the scene around it: "on a beach towel at golden hour," "held by a climber at a summit."

Use the same references every time. This is the single highest-leverage habit. A campaign generated against one fixed set of product references holds together; a campaign where each image re-uploads a slightly different crop does not.

Pin the style in reusable words

References pin the subject. Style needs a different tool: a short, fixed phrase bank that travels with every prompt. Write one sentence that names your look ("clean studio photography, soft directional light, warm neutrals, shallow depth of field") and paste it into every generation, unchanged. Resist the urge to embellish it per image. When the sentence changes, the look changes.

Model choice is part of style too. Each model has a visual accent: some lean editorial, some photoreal, some illustrative. Pick the one whose accent matches your brand and stay on it for a given campaign. Switching models mid-series is the fastest way to break a look, even with identical prompts.

Name your assets once

If your team generates through an assistant, references get even easier. With Dolly's MCP connector, uploads can be saved as named assets: store the product shots once as @product and the style frame as @look, and anyone (or any agent) can generate with "use @product and @look" from then on. No re-uploading, no wondering which crop a teammate used, no drift between the images you make and the ones your agent makes overnight. The Claude walkthrough shows this in practice.

A review pass that catches drift

Even with references and a fixed style sentence, review side by side. Drop new images next to three approved older ones before publishing. Drift is obvious in a lineup and invisible one image at a time. The cheap iteration helps here: at 5 to 15 credits per image, regenerating an off-brand result costs cents. Budget details are in the costs post.

Where this pays off

The payoff compounds when images feed other media. An on-brand still becomes the first frame of an on-brand video (see the product video walkthrough), which carries the same palette and product into motion. Get the stills right and everything downstream inherits it.