Mockupry

A mockup should not require Photoshop fluency or a marketplace subscription. Most still do.

Producing product mockups means either Photoshop skill or rigid template marketplaces. Both are slow. Both are a wall for non-designers who simply need a visual to ship.

The job is the same regardless of input. The input is just whatever the person already has.

I refused to pick one technique. A PSD template, a text description, a visual reference, a photo of the real world — each is a legitimate starting point, and forcing them all through one method would compromise all of them. So Mockupry runs three engines under a single workflow rather than one model behind a single button.

The three engines map to three signal colors in-product: PSD blue, Vision emerald, Generative violet. The color is the mental model. You always know which path produced your result and what its guarantees are. A PSD render preserves layers exactly; a vision composite respects real geometry and lighting; a generative pass invents. Those are different promises, so they read as different things.

The second decision was operational honesty. Heavy ML carries heavy dependencies, and I did not want the API's health tied to a GPU being present. The system is built to boot and degrade gracefully, then do the expensive work asynchronously where it belongs.

Three engines sit under one workflow: a PSD-native smart-object engine (psd-tools/pytoshop, layer-preserved export), an AI-vision engine, and a generative engine. The vision pipeline chains Grounding DINO for detection, SAM for segmentation, and Depth Anything for perspective, with dedicated geometry, lighting, occlusion, and mask-refinement modules. The generative pipeline uses pluggable model clients — Flux, Ideogram, GPT-Image, nano-banana — routed through fal.ai, Replicate, Together, and OpenAI, so a model choice is a config detail, not a rewrite.

The backend is async end-to-end: SQLAlchemy 2 with asyncpg, Celery and Redis for the heavy render and export jobs, and a WebSocket API for live job updates. The heavy model SDKs are lazy-imported, so the FastAPI app boots and degrades gracefully without GPU deps present. The frontend is Next.js 16 / React 19 / Tailwind v4 with a Fabric.js canvas editor and a tokenized per-engine color system. Storage runs on MinIO/S3, schema is Alembic-versioned, and Sentry handles observability.

In active development. The architecture — three engines, async job pipeline, provider-agnostic model clients — is in place; the work now is depth and breadth on each engine rather than reframing the platform.