SCS Studio — 3D Picture → IFC Modeling

X-ray of a populated 8-storey office tower — 805 pieces placed by the CP-SAT solver under German ASR workplace law, furniture visible through the ghosted building shell.

A main, fully finished project by Dimitres Kisimov. One photo in → a furnished, German-workplace-law-compliant BIM building out — shipped through four tagged releases (v1 pipeline → v2 editing & safety → v3 room parity → v4 procurement & ROI), every claim test-gated, downloadable below.

Photo-to-BIM pipeline: converts a single 2D photograph into a watertight, real-dimensioned, BIM-classified 3D object — then composes those objects into ergonomically furnished rooms and whole buildings, exported as optimized IFC4.

One app (npm start → http://localhost:3000, or double-click SCS_Studio.bat), three workspaces sharing one 3D viewport, plus a full Research & Comparisons hub (/hub.html, 9 guided steps) documenting every experiment behind it.

📦 Just want to run it? Grab the ~512 MB *_lite.zip from Releases, unzip, double-click SCS_Studio.bat — it self-installs on first run.

✨ Feature highlights across the four releases (click to expand)

• Photo → 3D: five AI engines benchmarked on 187 photos (TripoSR runs locally); broken meshes repaired with parametric chair-base grafts; every object auto-registered with a professional code (desk-TSG-042).
• Rooms: 18 selectable room types (offices → kitchens → server rooms, DE/EN), engine suggestions at three densities, collision checker with 0.90 m walking paths, honest refusals.
• Buildings: drop ANY IFC — 1,506 rooms across the 15-building demo fleet auto-classified (DIN 277, four languages) and furnished in minutes; floor-by-floor 2D/3D navigation.
• Editing everywhere: drag with live legality, 90° yaw rotation (R / ⟳ / Ctrl+right-click), delete-everywhere, lock/unlock — all baked into the exported IFC.
• German ASR compliance, cited: staffing caps (§5(3) ArbStättV), movement areas, route widths, protected access to fire extinguisher & first aid — mapping in docs/ASR_COMPLIANCE.md.
• 🛒 Procurement (v4): pick or drop any generated object → multi-company market sweep (IKEA · OTTO · POCO live; eBay & Google Shopping via keys) → CLIP visual matching → cheapest visually-similar product landed in Heilbronn, three business tiers, finance Excel pack. Honest claim printed on every report. docs/PROCUREMENT_METHOD.md
• 📊 ROI: measured, downloadable — one room 59% time saved, a 6-storey office 95.3%, the whole fleet 95.9% (docs/roi/).
• Proof: e2e 8/8 · floor dissection 15/15 · ergonomics 40/40 ×2 · exact-geometry meter over ~6,000 pieces · boot-verified release bundles.
• Open research, documented: look-alike product retrieval proved genuinely hard — three iterations are honestly graded in docs/PROCUREMENT_FINDINGS.md with a further-development roadmap. Negative results are results.

About this project

Built end-to-end by Dimitres Kisimov as a complete engineering effort — from research question to shipped, versioned releases. What it demonstrates in practice:

• Full-stack engineering: Node/Express backend, vanilla-JS 3D frontend (xeokit, model-viewer), Python compute layer — one app, one origin, no framework bloat.
• Applied ML with honest evaluation: five image-to-3D engines benchmarked on 187 photos against ground-truth meshes (F-score), CLIP/DINOv2 embeddings in production use, and a written record of what did not work.
• Operations research: CP-SAT constraint solving for furniture placement under real legal constraints (German ASR workplace law, cited from the legal text).
• BIM/AEC domain depth: IFC4 round-trips that survive Revit, DIN 277 room classification in four languages, 15-building validation fleet.
• Engineering discipline: every release test-gated (e2e, geometry meters, ergonomics checks), reproducible artifact scripts, honest limitation docs, clean licensing (ABO CC-BY-4.0, PolyHaven CC0, EU-compliant model selection).

---

Quick start — what needs to happen for this to work

Prerequisites: Node.js ≥ 18 · Python 3.11–3.14 · (optional) NVIDIA GPU, driver ≥ 520. Everything below is required once; afterwards npm start is the only command.

# 1. clone + Node dependencies (Express, xeokit)
git clone https://github.com/Dimitres-Kisimov/3DpicToIFCModeling.git
cd 3DpicToIFCModeling
npm install

# 2. Python dependencies
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126   # GPU
#   (no NVIDIA GPU? drop the --index-url flag — everything runs CPU-only, slower)
pip install -r requirements.txt
pip install transformers ultralytics "rembg[cpu]" ifcopenshell huggingface_hub shapely ortools

# 3. runtime config — defaults work out of the box
cp .env.example .env        # PORT=3000 · USE_GPU=true|false · PYTHON_PATH=python

# 4. ABO furniture library (retrieval catalog for "Build a room") — one-time, ~4.7 GB
python backend/python-scripts/download_abo_subset.py
python backend/python-scripts/build_abo_index.py       # DINOv2 + FAISS retrieval index

# 5. run
npm start                   # → http://localhost:3000

Included in the repo (nothing to fetch): the 605-item engine-generated catalog (data/generated_assets/), all six demo buildings (sample_buildings/, data/buildings/), the AI asset library, the full benchmark evidence (benchmark/), and the research hub.

Downloaded automatically on first use (cached by HuggingFace):

CPU-only fallback: with USE_GPU=false, TripoSR runs on CPU (~10–20 min per image instead of ~1–3 min). Room and building population are pure CPU by design — no GPU ever.

---

Research & Comparisons hub — /hub.html

Every experiment, benchmark and before/after, served from the same app:

• 🌐 Interactive 3D Building Explorer — all six populated buildings in live 3D (face-view buttons + NavCube, colored real-dimensioned furniture, 0 clashes, exact-geometry verified)
• 🧾 A/B lists 01–11 — 187 internet photos × every engine, with renders, watertightness, IoU stats and deep links into the spinning Multi-AI Visualizer (998 variants, engine emblems, 90° rotate, winner voting)
• 🏆 5-model gallery — TripoSG · SAM 3D · TRELLIS · InstantMesh · TripoSR on identical inputs (H200 study; A100 re-run reproduces F-scores within 0.003)
• 📖 Engine manuals — battle-tested install recipes, fix tables and lessons learned for 13 engines (deliverable/manuals/)
• 🪑 chair-graft, smoothing, leg-size and IFC-optimizer before/afters; system test report

Offline export: deliverable/research_export.zip (281 MB) packages the entire hub for people without GitHub — unzip, double-click START_WINDOWS.bat, done. Rebuild it anytime by re-running the staging steps in the export README.

Engine benchmark (Study C, A100 80GB, 187 photos — F-score ↑)

Model selection criteria, the HuggingFace census (625 → 11 EU-usable) and every elimination are documented in docs/MODEL_REQUIREMENTS_AND_ELIMINATIONS.md and docs/HF_CENSUS_2026-07.md. Nothing enters the catalog without passing repair → saveIFC → IFC4 validation; per-engine IFC evidence lives in benchmark/ifc/.

---

Building-scale population

populate_building.py reads every IfcSpace (name, true footprint, floor level), extracts the obstacles intruding into each room (walls, beams, columns, stairs — z-filtered to that room's own storey) plus door keep-clear zones, and runs the CP-SAT ergonomic solver (spatial_layout.py + rule_packs.py; ASR A1.2/A1.8 (Arbeitsstättenrichtlinie) for offices — default, Neufert / Panero-Zelnik / ADA clearances, circulation, no-overlap) to place furniture around them.

Robustness earned on real files:

• Rotated buildings — the rue-Marc-Antoine export models every wall at 60.4° to the world axes; the solver detects the rotation, solves in a de-rotated frame and rotates placements back.
• Non-rectangular rooms — L-shapes and internal voids are blocked via the space's true footprint polygon, not its bounding box.
• Colored, real-dimensioned furniture — per-material parts with baked PBR colors (xeokit renders no vertex colors), every asset normalised to its category's real Neufert dimensions.
• Honest reporting — items that don't fit are listed per room (dropped), never forced.

python backend/python-scripts/populate_building.py sample_buildings/Duplex_Architecture.ifc outputs/duplex_populated.glb

Verified across all six bundled buildings (exact polygon-intersection checks, not just bounding boxes): Duplex 8 rooms / 34 pieces, Schependomlaan (4 storeys) 34 rooms / 76 pieces, Kleine Wohnung 10 rooms / 30 pieces, COPROPIETE 16 pieces — 0 real clashes everywhere.

---

Stack

---

Project structure

3DpicToIFCModeling/
├── backend/
│   ├── server.js                  # Express entry point + static mounts
│   ├── routes/                    # /api: upload, generate, rooms, buildings, export
│   ├── services/pythonBridge.js   # spawns Python, parses JSON
│   ├── python-scripts/
│   │   ├── run_triposr.py         # photo → mesh pipeline
│   │   ├── repair_mesh.py …       # 7-archetype, 9-stage repair packs
│   │   ├── spatial_layout.py      # CP-SAT room solver
│   │   ├── rule_packs.py          # Neufert/Panero/ADA ergonomics
│   │   ├── populate_building.py   # whole-building population (CPU)
│   │   └── saveIFC.py             # IFC4 export + validation gate
│   └── triposr/                   # TripoSR source (MIT), marching-cubes patched
├── frontend/                      # SCS Studio + research hub + viewers/explorer
├── benchmark/                     # 11 A/B lists, visualizer, results, per-AI IFC evidence
├── data/
│   ├── generated_assets/          # 605-item engine-badged catalog (committed)
│   ├── buildings/                 # uploaded building IFCs (committed)
│   └── mesh_library_abo/          # ABO retrieval library (built locally, step 4)
├── deliverable/                   # manuals, cloud bundle, asset library, research export
├── docs/                          # user guide, security & compliance, engineering record …
├── sample_buildings/              # bundled demo IFC
├── package.json · requirements.txt · .env.example

---

API endpoints

---

Documentation

---

Known issues

• Python 3.14: torchmcubes has no wheels — patched with skimage.measure.marching_cubes in backend/triposr/tsr/models/isosurface.py
• xeokit vertex colors: the GLTF loader ignores COLOR_0 — all colors are baked as GLTF PBR baseColorFactor materials (this is handled automatically everywhere)
• Geometry-only engine exports (TRELLIS et al.) carry no textures; catalog items from them get realistic per-category material tones instead
• TripoSR orientation: upside-down output is corrected by a Y-centroid heuristic; unusual camera angles may still need manual rotation

---

Licenses & compliance

Commercial use: xeokit SDK and YOLOv8 are AGPL-3.0 — closed-source commercial deployment requires commercial licences from xeokit.io and Ultralytics. Territory compliance: Hunyuan3D-family models are excluded entirely (EU territory restriction — no research carve-out). Full audit: docs/SECURITY_COMPLIANCE.md.