OpenHouse 3D rebuilds a house as a 3D model from its real-estate listing photos, then renders a cinematic walkthrough video of it with headless Blender. You get photo-matched stills and a continuous fly-through/walk-through film. A comparison video plays the film above the original listing photos.

▶ Watch the side-by-side walkthrough (65 s, 3856 × 1980, 184 MB) · 1080p version (27 MB) · Project page

13695 Stanford Drive, Carmel, Indiana, rebuilt from its 32 listing photos. Left, Opus 5.5: the Stanford reconstruction in this repository, built with Claude Opus 5.5 using this toolkit. Right, Astra: an independent reconstruction of the same listing. Below each film are the original listing photo and the matching virtually staged render. The films are rendered from the 3D models; listing photos appear only in the comparison panels.

Under the hood:

• Camera solving from photos. Point and line solvers recover each listing photo's camera. They handle perspective-corrected (shifted-lens) photos and drone aerials, and solve plan dimensions jointly across photos.
• Procedural architecture in bpy. Walls with validated openings, lap siding and brick as real geometry, windows, doors, stairs, roofs with gutters, kitchens, baths and fixtures.
• Virtual staging and landscape. Furniture, soft goods, rugs, art rectified from the photos, trees, shrubs, grass, pavers, water and the neighbouring context.
• Production rendering. Headless Cycles on CPU, Metal, OptiX, CUDA, HIP or oneAPI; packed and fingerprinted scenes; resumable chunked rendering; motion-compensated temporal denoising; camera-route audits for clearance and whip pans.
• Delivery tools. Title cards, music edit, an offline photo/render gallery, frame-accurate comparison videos, and checksum inventories of the final files.

This is a modeling toolkit, not a one-click photo-to-3D service and not photogrammetry. A person or an AI coding agent reads the photos, sets out the plan and writes a small house package. The library does the geometry, rendering and video. Hidden geometry is inferred; a reconstruction is not a measured survey.

Requirements: Blender 5.2.x (tested with 5.2.1), Python 3.9+, and Git. Blender includes bpy; install the Python dependencies into a separate virtual environment. imageio-ffmpeg supplies FFmpeg for media tools. No API key is needed.

git clone https://github.com/yunfanye/openhouse-3d.git
cd openhouse-3d
python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev]'
export ARCHVIZ_PYTHON="$PWD/.venv/bin/python"

python tools/new_house.py my_house --title "My house"
blender -b --python-exit-code 1 --python run.py -- \
  --house my_house --cam hero --samples 16 --scale 25 --no-polish --device CPU

Open houses/my_house/output/renders/hero.png. The starter is a synthetic house that needs no photos, downloaded textures, or saved .blend file. Replace its plan, geometry, cameras, and materials with your house. The environment commands above use a POSIX shell; see setup for Windows and GPU selection.

1. Inventory your photos and their rights; write measured dimensions and uncertain assumptions in houses/my_house/REFERENCES.md.
2. Define one coordinate system, room volumes, levels, walls, and openings in plan.py. Solve photo cameras (tools/photo_grid.py, solve_camera.py, solve_scene.py, intcam_solve.py) before refining finishes.
3. Build architecture and permanent fixtures. Compare silhouettes, apertures, roof coverage, room connections, and fixtures against every available view (tools/overlay.py).
4. Add physically scaled materials, furnishing, vegetation, and lighting. Keep staging choices distinct from observed architecture.
5. Audit the baked camera route, inspect representative final-quality stills, freeze and pack the scene, then render into a fresh fingerprinted run.
6. Build the gallery and comparison video, decode the entire movie, inspect native frames and transitions, and record checksums before removing intermediates.

The detailed instructions are in the workflow, the realism checklist, and the production guide. Comparison manifests explain the gallery and video tools. Lessons record what went wrong on each house and how it was fixed.

Keep transferable algorithms in archviz/; keep dimensions, photo mappings, art direction, and architectural decisions in houses/<name>/. The library does not import a specific house. See architecture.

The listing photos for Alpine, Stanford, Walsh and Webster are included so the examples can be reproduced: 131 reference images, including Webster's floor plan, and the existing source manifests. Image bytes are preserved; Stanford's filenames are normalized to 01.jpg through 32.jpg. See photo removal requests if an image should be removed.

Generated frames, models, movies and local environments are ignored. New house photo folders stay local by default. Final media are published as release downloads.

python -m pytest -q
python -m ruff check .
python -m archviz.delivery houses/my_house/output --record
python -m archviz.delivery houses/my_house/output           # verify + list cleanup candidates
python -m archviz.delivery houses/my_house/output --prune   # keep only verified final/

Put accepted artifacts in output/final/ before recording the inventory. The checksum file is a record of accepted bytes, not proof that a model is accurate. Never replace visual review with a green test suite.

Project code and documentation are MIT licensed. Reference photographs, listing floor plans, third-party assets, and media incorporating them retain their own rights; the code license does not relicense those files. See asset provenance and publishing.

Contributions should include a small reproducible example and evidence appropriate to the change. Start with CONTRIBUTING.md.