the-long-take

Twenty-four hours of one room, rendered as a single continuous translucent water-membrane you can listen to. Built for haroldathome.com/the-long-take; released for anyone with a Home Assistant install and an mmWave radar.

A long take is a film term — Hitchcock made Rope (1948) as one apparent continuous shot. This is a long take of one room, fixed camera, twenty-four hours, no cuts. The room plays.

The radar reports body positions eight times a second. The pipeline pulls the day's positions from your Home Assistant database, decimates them, and writes a small JSON. The renderer reads the JSON and draws a translucent membrane that bulges where bodies were, thins where the room was empty, and grows along a horizontal time axis. A pentatonic synth chimes when the playback cursor crosses each thirty-second bin. Audio is opt-in.

---

At a glance

• Renderer: Three.js + WebGL2, water-correct PBR, fbm-perturbed normals, UnrealBloom post-processing, Web Audio synth with synthetic convolution reverb. ~1100 LOC, single file.
• Data pipeline: Python + pymysql + numpy. Pulls LD2450 X/Y/speed columns from the HA MariaDB recorder, pairs by timestamp, decimates, ships JSON to the renderer host.
• Schedule: 5-minute systemd timer. The day on your wall is always within 5 minutes of live.
• Format: one JSON per room, ~200–700 KB after decimation.
• Browser audience: modern Chromium, Safari, Firefox. Requires WebGL2.

---

Quick look — no install required

git clone https://github.com/alfiedennen/the-long-take.git
cd the-long-take

# Browsers block file:// fetches across directories, so a tiny local
# server is required to load the bundled sample.
python -m http.server 8000

Then open http://localhost:8000/examples/?r=office. You should see a translucent membrane, slowly orbiting, with a scrubber along the bottom. Tap TAP TO ENABLE SOUND in the top-right; press space to play; drag to orbit; scroll to zoom.

The bundled sample is one anonymous day in one office — about 30,000 radar events, decimated from ~58,000 originals.

---

Run it on your own house

What you need

• Home Assistant 2024.6+, with the MariaDB recorder backend (the default SQLite recorder hits performance cliffs at the LD2450 scan rate — switching to MariaDB is a single-line config change but takes some setup).
• One Hi-Link LD2450 mmWave radar per room you want to render. Other presence sensors (PIR, FP300, mmWave-without-position) won't work — the long take needs per-target X/Y position over time.
• Python 3.11+ with pymysql and numpy.
• A small always-on Linux machine that can SSH to your Home Assistant host (any LXC, Pi, or mini-PC). The pre-compute runs here, not on HA itself.

Setup, in five steps

1. Create a read-only MariaDB user dedicated to this script. On your MariaDB host:

   CREATE USER 'the_long_take_reader'@'localhost' IDENTIFIED BY 'pick-a-strong-password';
   GRANT SELECT ON homeassistant.* TO 'the_long_take_reader'@'localhost';
   FLUSH PRIVILEGES;

2. Install the precompute on your Linux host:

   git clone https://github.com/alfiedennen/the-long-take.git
   cd the-long-take/precompute
   python3 -m venv .venv && source .venv/bin/activate
   pip install -r requirements.txt
   cp rooms.example.py rooms.py

   Edit rooms.py to map your LD2450 entity names. (See docs/adapt.md for the entity-naming details.)

3. Configure the environment:

   export EXPOSURE_DB_PASSWORD='your-strong-password'
   export EXPOSURE_HA_HOST='root@homeassistant.local'

4. Run it once manually to verify everything works:

   python exposure_precompute.py 24

   You should see point counts per room printed, then Deployed to root@homeassistant.local:/config/www/the-long-take/data/.

5. Install the systemd timer for the 5-minute refresh. See docs/install.md for the full systemd setup.

Mount the renderer in Home Assistant

scp -r renderer/ root@homeassistant.local:/config/www/the-long-take

Visit https://<your-ha>/local/the-long-take/?r=office for one room. (Replace office with whatever room_id you used in rooms.py.)

---

How it works

For the architectural deep-dive — the binning strategy, the membrane shader, the audio synth, the eight pitfalls captured during v22 — see docs/how-it-works.md.

---

Adapt to your setup

For mapping LD2450 entity names, adjusting room dimensions, changing the visual lock, sub-day windows, single-room vs multi-room — see docs/adapt.md.

---

Known limits

• LD2450 only. Requires per-target X/Y position. PIR, FP300 (Aqara), and other binary presence sensors won't work.
• MariaDB only. SQLite recorder won't scale to LD2450's 8 readings/sec across multiple rooms over a 24h window.
• One day default. Longer windows (a week, a month) need decimation re-tuning — the current TARGET_POINTS = 30_000 is calibrated for 24h.
• Audio requires user gesture. Browser policy. The first click anywhere enables sound.
• WebGL2 required. No fallback for older browsers.
• Single-room per render. Each iframe / browser tab renders one room. The piece's grandeur is in viewing each room's tube alongside the others — but the renderer doesn't do this internally; it's a deployment / dashboard composition concern.

---

Lock state

• Renderer: v22 (10 May 2026)
• Pre-compute: v1 (8 May 2026)
• Visual journey, design lock state, and the eight captured pitfalls live on haroldathome.com/the-long-take.

See CHANGELOG.md for release notes.

---

Credits

• Original work at haroldathome.com by Alfie Dennen, Hastings, England.
• Built with Three.js, pymysql, numpy.
• Inspired by Hitchcock's Rope, BERG, Holo, LIGO data viz, and the fixed-camera tradition.
• Thanks to the Home Assistant community for MariaDB recorder hardening notes and the LD2450 ESPHome integration.

---

Licences

Scope	Licence
Code (renderer/, precompute/, examples/index.html, .githooks/)	MIT — see LICENSE
Documentation, sample data, prose, screenshots	CC-BY-NC 4.0 — see LICENSE-CONTENT

---

See also

For the architectural context this piece sits in — the wider house system, the AI agent named after the previous occupant, the wall display, the wake words, the painted axonometrics that change with who is home — see haroldathome.com.