What it does 3DCellForge is a React + Three.js workbench for uploading reference images or GLB files, generating 3D models through external providers, then inspecting, comparing, and presenting the results. It keeps a local model library in IndexedDB, exports screenshots and GLBs, and includes a demo mode with cinematic camera paths tailored to asset type—road push-in for cars, flight pass for aircraft, studio orbit for organic specimens.

The interesting bit The project treats multiple competing image-to-3D services as interchangeable backends. It supports Hyper3D Rodin, Tripo, Fal.ai (with model selection across Hunyuan3D v2, TRELLIS, TripoSR, and others), local Hunyuan3D, and even a browser-side JS Depth fallback—then caches generated GLBs locally so you aren’t burning API credits on every reload. Optional OpenAI vision classification improves prompts when configured.

Key highlights

• Six generation modes including cloud APIs, local server, and a layered PNG depth fallback
• IndexedDB persistence with localStorage fallback; models survive refresh
• Quality scoring on generated GLBs: triangle count, texture count, file size, demo readiness
• Demo mode hides chrome and runs asset-aware cinematic camera animations
• Playwright visual regression tests for workbench, library drawer, and demo mode
• API keys stay server-side in .env.local; frontend never sees them

Caveats

• Requires Node backend running separately (npm run dev:api) for generation features; frontend alone only handles local GLB import
• Hunyuan3D local mode expects a specific API shape (POST /send, GET /status/:uid) that you must provide yourself
• Demo video thumbnail exists but the actual video is a local MP4, not a hosted stream

Verdict Worth a look if you’re building 3D asset pipelines and want a unified interface to compare multiple generative providers. Skip it if you need a finished product—this is a well-structured prototype with real provider integrations and thoughtful caching, but still a prototype.