Photowa
Non-designers often struggle to recreate high-quality marketing materials. They have a reference image but lack the technical skill to recreate the layout in tools like Photoshop or Figma. Photowa solves this by bridging the gap between visual inspiration and editable reality. By feeding a reference image to an AI vision model, the platform instantly reconstructs the layout as an interactive, fully layered canvas.
Timeline
4 Months
Services
Tech Stack
The Creative Bottleneck
Generative AI image models are great at producing beautiful pixels, but they fail utterly at producing editable layouts. If a marketer uses Midjourney to create a poster, they cannot easily change the text, move the logo, or swap the background.
We set out to build a system that didn't just 'generate an image', but actively reverse-engineered a layout into a structured, editable DOM.
Taming the Vision Model
Our first attempts at prompting GPT-4 Vision resulted in chaotic, hallucinated coordinates. The breakthrough came when we stopped asking the model to 'describe the layout' and instead forced it into a strict JSON schema using Pydantic.
By constraining the output to specific x, y, width, and height floats, we turned a subjective generative model into a deterministic layout extraction engine.
The Live Canvas
Extracting the coordinates was only half the battle. We built a robust Next.js frontend utilizing Fabric.js to instantly hydrate those coordinates into a live HTML5 Canvas.
Users can drag, drop, recolor, and edit the typography of any extracted element. The platform essentially gives non-designers the superpower to 'inspect element' on flat JPEGs.
Design Workflow
Eliminated the manual rebuild step, providing an instant editable live canvas.
Accuracy
Maintains exact layout through Pydantic-typed structured output.
User Base
Dramatically lower barrier for non-designers to reproduce professional layouts.
“Photowa bridges the gap between 'I like this layout' and 'ship my version of it' by giving me the layout itself—structured and editable—instead of just a picture.”
Users don't want a picture of a layout; they want the layout itself.
Typing AI vision output into schemas avoids the brittleness typical in 'AI-to-UI' generation attempts.
Asynchronous architectures are non-negotiable when combining LLMs, external APIs, and persistent databases.