Open House
Open House turns phone photos of a home into a walkable 3D tour with listing-ready golden-hour edits.
Core trick
Argon scores each generated golden-hour photo edit and sends weak ones back for redo, under validators and step checkpoints.
Description
Open House turns phone photos of a home into a walkable 3D tour with listing-ready golden-hour edits.
Gemini 4 Argon built the app in Google AI Studio from one brief, then ran the pipeline: ingest (photos → rooms, materials, objects), layout (rooms in metres, doors, furniture, one camera per photo), golden-hour edit plans, an automated judge that scores every edit and sends weak ones back for a redo, and narration. Everything runs inside a harness written beforehand and loaded from GitHub at a pinned commit: validators on every answer, a checkpoint after every step, resume from any step, and a report recounted only from the call trace.
Today's run (10 photos, 8 rooms): 41 Argon calls + 14 image edits, 1 validation retry, 4 edits redone (2 rose from 3/5 to 5/5, 2
README excerpt
# Open House — app Turn phone photos of a home into a walkable 3D tour with golden-hour edits. This app was built by **Gemini 4 Argon in Google AI Studio** (Build mode) at the Gemini 4 Argon SF Hackathon, Oct 25 2026. It is a thin UI: the pipeline it runs is the Open House harness, loaded from GitHub at a pinned commit — not copied into this repo: - `https://cdn.jsdelivr.net/gh/icomppower/open-house@3f38f72/src/…` (harness, Gemini adapter, 3D viewer, report, export) - Harness repo: https://github.com/icomppower/open-house **Live app:** https://open-house-ai.ai.studio · **Demo video:** https://youtu.be/hjP61BlFWkM ## What it does Upload photos → six steps run on Gemini 4 Argon (`gemini-4-argon`) with a checkpoint after each: prepare → ingest (rooms, materials, objects) → layout (rooms in metres, doors, furniture, one camera per photo) → enhance (golden-hour edit plans; edits by the image model) → judge (scores every edit and tour stop, redoes weak edits) → narrate. Every step's answer is checked by the harness validators; failed steps can be resumed; "How it was made" is rebuilt only from the call trace. "Load saved run" replays a saved run state with no model calls. Thinking levels used: `ingest: 'medium'`, `layout: 'low'`, `narrate: 'low'` (layout at `'high'` exceeded a 300-second per-call limit on 8+ rooms). ## Hackathon run (10 photos, run `run_mvo7ujv7_12cgat`) 55 calls (41 Argon + 14 image edits) · 1 validator retry · 4 edits redone (2 rose from 3/5 to 5/5, 2 fl