Google Patents an AI That Writes Its Own Image Prompts to Build Ads

By Patentlyze Team · Updated Jul 3, 2026

Google is patenting a system where AI not only generates ad images but also figures out what those images should look like in the first place, by writing its own creative prompts along the way.

How Google's AI turns a product into ad images automatically

Imagine you sell hiking boots, and instead of hiring a photographer or a designer, you just describe the boots to an AI and it figures out: 'These boots suggest rugged mountain trails at sunrise.' Then it goes off and generates a set of photorealistic images of exactly that scene, with your product's qualities baked into every shot.

That's essentially what Google is patenting here. The system takes a list of a product's features, asks a language model to dream up fitting visual settings, and then hands each of those settings off to a separate image-generation model to produce the actual pictures. The whole process is automatic, no human briefing a designer or writing a single prompt.

At the end, the AI picks one of those images and wraps it into a finished ad ready to deliver to a user's screen. For advertisers running large-scale campaigns with hundreds of products, this could replace a significant chunk of the creative production process.

How the two-step prompt chain generates and selects images

The patent describes a two-stage pipeline Google calls a 'prompt chain.' Rather than asking one AI model to go straight from product description to finished image, the system breaks the job into two handoffs.

Stage one: An AI system takes a set of attributes for an item (think: 'durable, waterproof, lightweight, designed for cold weather') and passes them to a language model (a text-based AI, like a GPT-style system). The language model's job is to output a list of environments that would visually communicate those attributes to a viewer, for example, 'a snowy alpine ridge at dusk' or 'a misty Pacific Northwest forest trail.'

Stage two: For each environment the language model proposes, the system automatically generates a new, more detailed prompt and feeds it to a separate image generation model (an AI that produces pictures, like a Stable Diffusion or Imagen-style system). That model produces one or more images for each setting.

Once the full set of images is ready, the system selects one, wraps it in a digital component (the patent's term for an ad unit or content card), and delivers it to a user's device. The key claim is that no human needs to write the image prompts manually; the AI chain handles that translation from product attributes to visual concept to finished image.

What this means for automated digital advertising

For Google's advertising business, this patent points toward a future where ad creatives are generated on the fly, personalized to a product's specific traits, and produced at a scale that would be impossible with human designers. If you run Google Ads today, this kind of system could eventually mean your campaigns generate their own visuals based on your product feed, skipping the step where you upload images.

For the broader ad industry, it's a signal that the creative layer of digital advertising is increasingly something AI will handle end-to-end. The patent doesn't just cover image generation; it covers the strategic reasoning about what kind of image to generate, which is traditionally a human creative decision. That's the part worth watching.

Editorial commentary on a publicly published patent application. Not legal advice.