Inspiration
What it does
How I built it
Challenges I ran into
Accomplishments that I'm proud of
What I learned
What's next for FurniMesh: AI 3D Furniture Generator
Inspiration
Furniture manufacturers and e-commerce devs spend hours manually modeling products from photos for catalogs and AR viewers. We wanted to automate the tedious from photo to 3D model pipeline with AI that understands furniture anatomy. Built as NextGenerationEU-funded research to bridge AI vision with professional 3D workflows.
What it does
FurniMesh takes single or multi-view furniture photos and generates production-ready 3D models using specialized AI trained on global trade fair datasets. It automatically segments components (legs, seats, backrests, drawers) into color-coded meshes with clean topology optimized for Blender, SketchUp, and Corona Renderer. Exports in GLB/OBJ/SKP/Blend formats; includes an AR-enabled web viewer for instant e-commerce embedding with one-tap mobile AR.
How we built it
Custom NeRF-based 3D reconstruction pipeline fine-tuned on 50k+ furniture images, with a novel furniture-specific segmentation model using SAM2 + domain adaptation. Backend in PyTorch/Docker on AWS/GCP; frontend React/Three.js for interactive preview and configurator. Integrated differentiable rendering for topology cleanup and UV unwrapping automation.
Challenges we ran into
Furniture's reflective surfaces and complex assemblies broke generic image-to-3D AI, requiring custom multi-view consistency losses and part hierarchy prediction. Balancing photorealistic textures with clean quad topology for pro workflows took 3 months of iterative training.
Accomplishments that we're proud of
Achieved 85% automatic part separation accuracy on real-world catalog photos (vs 40% for generic tools); models load 5x faster in SketchUp than manual imports. Powered 200+ e-commerce sites with AR viewers, cutting product modeling time from 4hrs to 4min.
What we learned
Domain-specific training data trumps scale—our 50k curated furniture images outperformed generic billion-scale models. Semantic part understanding is the real bottleneck in 3D reconstruction; geometry alone isn't enough for pro usability.
What's next for FurniMesh
API launch (May 2025) for headless batch generation + real-time configurator SDK. Multi-modal inputs (text+image) and physics-aware simulation for AR try-on experiences.
Log in or sign up for Devpost to join the conversation.