Picturesque: Unifying 50+ AI Models Under One Roof
Picturesque unifies 50+ AI models for images, video, audio and music under one credit system, covering the architecture choices and lessons learned along the way.
Picturesque is a creative studio that brings together more than 50 AI models for images, video, music, audio, editing, upscaling and motion control under a single interface and one shared credit balance. The team's real challenge wasn't model quality but the surrounding plumbing: every provider has its own UI, billing rules and input schema, forcing users to jump between tools to go from generating an image to animating it and adding music.
On the engineering side, the stack was kept deliberately boring: React/TypeScript/Vite on the frontend, Node/Express on the backend, Socket.IO for real-time updates, Supabase for data and auth, and FFmpeg/Sharp for media processing. The hard part was building a normalization layer that unifies dozens of models under one pricing engine, one generation flow and one history table — including a painful but worthwhile refactor of a 7,700-line server.js file into proper routes and services.
The product itself spans Studio tabs for manual control, a node-based Workflows canvas for chaining models, and Director, an agent layer that plans and executes multi-step creative tasks using the same underlying generation code. The write-up shares hard-earned lessons around instant refunds on failure, race conditions in real-time updates, and the ongoing maintenance burden of keeping a model config layer current — practical takeaways for any engineer building a multi-provider generative AI product.