InteriorFusion: Single Image → High-Quality Editable 3D Interior Scenes

InteriorFusion is a production-grade, open-source AI system for converting a single 2D interior photograph into a high-quality, editable 3D scene with textured meshes, Gaussian splats, and PBR materials.

🏆 Why InteriorFusion?

Current image-to-3D pipelines (TRELLIS, Hunyuan3D-2, TripoSR, Stable Fast 3D) are designed for single objects, not rooms. InteriorFusion is the first system specifically architected for interior design and room reconstruction.

Key Differentiators

• Scene-aware: Understands room layouts, wall/floor topology, and spatial relationships
• Multi-object coherent: Furniture doesn't float; objects respect real-world scale
• Editable: Individual furniture objects are separable, movable, replaceable
• PBR materials: Metallic, roughness, normal maps for photorealistic rendering
• Multi-format export: GLB, FBX, OBJ, USDZ, Gaussian Splatting
• Under 30s generation on RTX 4090

Architecture Overview

Single Interior Image
       │
       ▼
┌─────────────────────────────────────┐
│ Phase 1: Scene Understanding        │
│ - Depth Anything V2 (metric indoor) │
│ - Room layout estimation             │
│ - Semantic segmentation             │
└─────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────┐
│ Phase 2: Multi-View Generation       │
│ - Zero123++ / SyncDreamer            │
│ - Depth-conditioned view synthesis   │
└─────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────┐
│ Phase 3: 3D Reconstruction           │
│ - Room shell (walls/floor/ceiling)    │
│ - Per-object Gaussian reconstruction │
│ - TRELLIS.2 for furniture detail    │
└─────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────┐
│ Phase 4: Scene Assembly             │
│ - SpatialLM layout optimization      │
│ - Collision detection & correction   │
│ - Scale normalization               │
└─────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────┐
│ Phase 5: Material & Texture         │
│ - PBR material generation            │
│ - Texture baking                     │
│ - Light estimation                   │
└─────────────────────────────────────┘
       │
       ▼
  Export: GLB / FBX / OBJ / USDZ / 3DGS

Model Zoo

Model	HF Path	Purpose	Size
InteriorFusion-S	stevee00/InteriorFusion-S	Fast preview, 512³	1.5B
InteriorFusion-L	stevee00/InteriorFusion-L	Production quality, 1024³	4B
InteriorFusion-XL	stevee00/InteriorFusion-XL	Research quality, 2048³	10B

Quick Start

pip install -e .
python -m interiorfusion.infer --image room_photo.jpg --output ./output/

License: MIT