Imagen over Stock Photos

Instead of licensing generic stock photography, I generate all visual assets for Total Ventures' products using models like Imagen, ensuring every image is contextually relevant and unique to the brand.

What it is

Imagen over Stock Photos means I generate custom, contextually relevant images using generative AI models at the point of need, rather than relying on pre-existing stock photo libraries. This ensures that every visual asset is unique to the brand and specific content it supports. For Total Ventures, this is less about creating photorealistic images and more about generating illustrations, abstract visuals, or specific product-related graphics that align precisely with a product's aesthetic and message. It's a deliberate choice to move away from the generic visual language that often comes with licensed stock assets.

Why I do it this way

My primary motivation is brand consistency and uniqueness. Stock photos, by their nature, are designed to be broadly applicable, which often makes them generic. For a multi-product studio like Total Ventures, each brand — whether it's Total Formula 1, Pregnancy Power Hour, or Inky — needs a distinct visual identity. Generative AI gives me precise control over style, subject matter, and mood. This approach is a natural extension of my Agent-First Development philosophy, where autonomous agents handle more of the creative and operational load. It eliminates recurring licensing fees, which, while not massive individually, add up across multiple products and content pieces. More importantly, it provides speed and agility; I can generate images on demand, integrated directly into content workflows, rather than spending time searching, licensing, and editing stock. This supports rapid iteration and shipping.

How it works in practice

My process begins with developing robust prompt templates for different image needs. For a blog header on Pregnancy Power Hour, for example, the prompt might specify a watercolor illustration style, a serene mood, and a specific subject like 'a pregnant woman doing yoga, gentle light.' Often, a content agent, orchestrated by what I call Claude Code as the Engineer, will generate not only the text for a blog post but also the specific prompt for its accompanying image. This prompt is then sent via API to a generative model like Imagen. The output is reviewed, and if necessary, the prompt is refined for a new generation. This feedback loop is often automated for initial passes. Once an image is approved, it's stored in Firebase Storage and served via Vercel's CDN for optimal performance. For Inky, this means generating custom, abstract illustrations for product features or UI elements that would be impossible to find in a stock library. For Total Formula 1, it might involve stylized graphics depicting race dynamics or historical moments, maintaining a consistent visual language across all content.

Where this breaks down

While effective, this approach has its limitations. Generative models can struggle with fine details, accurate text within images, or specific anatomical precision, which is critical for medical content on Pregnancy Power Hour. This requires careful human oversight and often multiple prompt iterations. Bias inherited from training data is another concern, necessitating vigilant review and prompt engineering to avoid unintended representations. The computational cost of API calls, while often less than licensing, can accumulate, especially during heavy content generation phases. Finally, developing effective prompt engineering skills and integrating these workflows takes time. Maintaining a consistent visual style across many generations and products can be challenging without careful prompt management, often requiring a detailed style guide documented within my Docs as Source system to ensure agents adhere to brand guidelines.