3D Room Project · Handoff

Recreate a room in 3D — project handoff

This page is the bridge between two laptops. Machine A (integrated GPU) built a hand-built walkable model. Machine B (RTX 5060) will build the photoreal Gaussian Splat from a folder of room photos. Everything below is the full set of instructions — the photos live on the 5060 laptop (not hosted here for privacy).

🤖 If you are Claude (or any AI agent) reading this

The user said something like "check 3droom.sitemodernize.com". Your job: help them build a 3D Gaussian Splat of their room from a folder of photos on this RTX 5060 laptop, then view it in three.js. Do this:

Find the photos locally. The user has a folder of room photos (JPG/PNG) on this laptop. Ask them for the folder path (e.g. photos\). The earlier home video is NOT used and NOT hosted here — input is the new photos.
Get the viewer. Download splat_viewer.html from this site into the working folder:
```
curl -O https://3droom.sitemodernize.com/splat_viewer.html
```
Train the splat. Follow Track A (Postshot) below — easiest and Blackwell-safe. Feed it the photos folder. Brush / Nerfstudio are alternatives. Output a file named room.ply.
View it. Put room.ply next to splat_viewer.html, run python -m http.server 8099, open http://localhost:8099/splat_viewer.html. Fix orientation knobs at the top of that file if the scene loads tilted/upside-down.

Hardware: RTX 5060 laptop, 8 GB VRAM, Blackwell (sm_120). Prefer tools that support CUDA 12.8+ or run on wgpu. Good capture = locked exposure/focus/WB, 150–400 overlapping stills, translate don't just rotate, multiple heights, keep furniture/edges in frame (blank white walls are hard for SfM).

Background — what's already done

On the first laptop we built a hand-built, walkable three.js model of the apartment (living room + dining room + hallway): first-person WASD controls, collisions, procedural wood floor, real family-photo textures on the picture frames, glossy PBR materials, image-based lighting, ambient occlusion + bloom, and a Low/Med/High quality switch (press P). It's stylized, not photoreal — that's why we're now doing the splat.

That model and its personal textures are not published here. This handoff is about the photoreal path.

Goal

Turn a folder of room photos into a 3D Gaussian Splat (photoreal, real pixels), trained locally on the RTX 5060 (no cloud upload), and view it in the browser with splat_viewer.html. Photos beat video on quality — full resolution, no motion blur or codec artifacts.

Shooting the photos

Lock exposure + focus + white balance (phone Pro/manual mode).
150–400 stills, 60–80% overlap; translate, don't just rotate (splats need parallax).
Multiple heights, every corner, orbit the furniture; keep furniture/edges in frame (blank white walls are hard for SfM).
Put them all in one folder, e.g. photos\.

Pick a training tool

Tool	Ease	Blackwell (RTX 50)	Cost	Built-in SfM?
Postshot (recommended)	GUI, photos in → splat out	✓ supports RTX 50	free trial / paid	yes
Brush	open source, no CUDA (wgpu)	✓ via DX12/Vulkan	free	no — needs COLMAP
Nerfstudio splatfacto	CLI / Python	⚠ needs torch cu128 nightly	free	no — needs COLMAP

Brand-new 50-series + CUDA can be painful. Postshot (all-in-one GUI) or Brush (no CUDA at all) avoid the toolchain headache. Start with Postshot.

Track A — Postshot (easiest, recommended)

Install Postshot: jawset.com
New Project → add the photos folder (drag the folder of images in).
Settings for 8 GB VRAM: profile Splat MCMC or Splat3 (defaults fine). If it OOMs: lower Max Splats (~1.5M) or downscale images to ~1600 px.
Start training (~5–20 min on a 5060 for one room).
Export → PLY (3D Gaussian Splat) → save as room.ply next to splat_viewer.html.

Track B — Brush (free, open source, no CUDA)

Trains via wgpu (DirectX 12), so Blackwell "just works." Needs camera poses (COLMAP) first, computed directly from the photos.

Install COLMAP: github.com/colmap/colmap/releases
COLMAP automatic reconstruction on the photos folder (GUI Reconstruction → Automatic, Image folder = photos; or CLI):
```
colmap automatic_reconstructor --workspace_path . --image_path photos --camera_model OPENCV
```
Download Brush: github.com/ArthurBrussee/brush/releases → open the COLMAP workspace → train → export room.ply.

Track C — Nerfstudio splatfacto (advanced)

Python only. Blackwell needs the CUDA 12.8 / PyTorch cu128 build:

pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install nerfstudio
ns-process-data images --data photos --output-dir room_proc   # runs COLMAP on the photos
ns-train splatfacto --data room_proc
ns-export gaussian-splat --load-config <config.yml> --output-dir export    # -> .ply

Rename the export to room.ply.

View it (any track)

Put room.ply in the same folder as splat_viewer.html.
Serve over http (ES modules + workers need it):
```
python -m http.server 8099
```
Open http://localhost:8099/splat_viewer.html
Mouse drag = orbit, scroll = zoom, WASD/QE = fly, R = reset.

If the room loads tilted/upside-down (common — axes depend on the solver), edit CAMERA_UP / INITIAL_LOOK_AT at the top of splat_viewer.html (try [0,1,0] vs [0,-1,0]).

Tips

8 GB VRAM: if training OOMs, downscale frames (1600 px wide is plenty) and cap splat count. One room fits easily.
More photos + more overlap = better splat. 200–400 sharp overlapping shots is a sweet spot.
White textureless walls are the hard part for SfM (any tool). Mitigate with heavy overlap and by keeping furniture/edges/corners in frame, not blank wall.
.ply can be 100–400 MB. For faster web loading, export compressed from Postshot or convert to .ksplat with the @mkkellogg/gaussian-splats-3d tool. Raw .ply works, just slower to load.

Files on this site:
splat_viewer.html RTX5060_SPLAT_GUIDE.md

Personal media (the room photos and any family-photo content) is intentionally not hosted here — it stays local on the 5060 laptop.