Imagenerator, an AI image and video generation studio

Imagenerator is Mediapilote’s internal AI image and video generation tool: garments shown on virtual models, free creation from reference images, short videos. The first version ran on Express and PocketBase.

In 2026 its use grew: more models, longer generations, print requirements and studios dedicated to some of the agency’s clients. The old stack could no longer keep up, so I rebuilt the platform on Cloudflare.

An image or video generation can take several minutes and fail on the provider’s side. Nothing could be lost: not the running generation if the user closes the tab, not the reference images, not the AI-enhanced prompts.

Designers also needed print-ready files, meaning layered PSDs and CMYK, which AI models do not produce. Finally, client studios had to share the same code while keeping their own identity and scope.

I replaced the home-made queue with Cloudflare Workflows: the server itself saves generations, which stay visible everywhere while they run, with up to 10 jobs in parallel. All models go through OpenRouter, with an AI judge that scores the results. I added mask-based targeted retouching, local background removal with BEN2 on WebGPU, an image upscaler and a Cloudflare container that converts PSDs to CMYK with an ICC profile.

For clients, I designed a base for modular studios connected to the intranet through SSO, with a product catalogue fed by web search and an AI-assisted newsletter editor. A personal MCP server, protected by OAuth, also lets users drive the studio from an AI assistant.