☄️ Photogram
Relive your memories in true depth. Organize and explore your visual history as fully rendered, locally-hosted 4D Gaussian Splats.
Overview
Photogram (Companion Photo Time Machine) is a locally-hosted library and viewer for 3D Gaussian Splatting — a state-of-the-art technique that reconstructs photorealistic 4D scenes from ordinary photo or video collections. Unlike flat photo galleries, Photogram lets you step inside your memories and experience them with depth and parallax.
All processing runs on your Hub, using your GPU. The resulting splat files are stored locally, integrated with Digital Memory, and viewable in any WebXR browser or on a VR headset.
Key Features
- 4D Gaussian Splatting — reconstruct volumetric scenes from photos or video
- Local training — all NeRF/3DGS training runs on your Hub GPU
- Photo library management — import, organize, and tag photo collections for reconstruction
- Timeline integration — splats appear in Digital Memory and Spatial Time Machine
- WebXR viewer — view splats in browser, in VR, or on mixed reality headsets
- Batch processing — queue multiple reconstructions to run overnight
- Export — export as
.splat,.ply, or video flythrough
What is Gaussian Splatting?
3D Gaussian Splatting represents a scene as millions of tiny ellipsoidal “splats,” each with color, opacity, and position parameters. Trained from a photo set, the result is a photorealistic, freeform 3D representation navigable from any viewpoint — far more lifelike than traditional photogrammetry meshes.
Hardware Requirements
| Component | Minimum | Recommended |
|---|---|---|
| GPU VRAM | 8 GB | 16 GB+ |
| System RAM | 16 GB | 32 GB+ |
| Storage | 50 GB free | 200 GB+ |
| CUDA (NVIDIA) or ROCm (AMD) | Required | — |
Training time scales with scene complexity and photo count. A 200-photo scene takes approximately 15–20 minutes on a 16GB GPU.
Use Cases
- Reconstruct a vacation, event, or milestone in photorealistic 3D from an album of photos
- Create a volumetric record of a location before it changes (construction, renovation, sale of a home)
- Build immersive memories to explore in the Spatial Time Machine
- Generate a walkaround visual reference of a physical space for design or insurance purposes
- Archive memories in a format that preserves spatial depth, not just pixels
Setup
Install from Hub
Search for Photogram in the Hub app store and install. The setup wizard checks GPU availability and CUDA/ROCm drivers.
Prepare your photo set
Collect 30–500 photos of the scene/subject you want to reconstruct. Tips:
- Shoot from a wide variety of angles and distances
- Ensure consistent lighting (avoid mixed indoor/outdoor in one set)
- Use sharp focus; motion blur degrades reconstruction quality
- Overlap between consecutive photos of at least 70%
Create a project
Open http://photogram.ci.localhost → New Project. Upload your photos or point to a folder on your Hub’s storage.
Run reconstruction
Select reconstruction method:
- Quick Splat — faster, lower detail, good for preview
- Full 3DGS — full quality, recommended for archival
Click Train and wait for completion. Monitor progress via the live loss graph.
Usage
Viewing a Splat
Click any completed project to open the 3D viewer. Use mouse/trackpad to orbit, and scroll to dolly in/out. For WebXR/VR:
- Click Enter XR in the top-right corner
- Put on your headset and use hand tracking or controllers to navigate
Annotating a Splat
While in the viewer, press N to add a floating note anchored to a specific point in the scene. Notes are synced to Digital Memory.
Exporting a Flythrough Video
Navigate to Project → Export → Video. Define a camera path by placing keyframes and render a 4K MP4 flythrough.
Batch Reconstruction
Queue multiple projects in Projects → Batch Queue. Set them to run at a specific time (e.g. overnight) to avoid impacting Hub performance during the day.
Troubleshooting
Training fails immediately
Check GPU driver compatibility. For NVIDIA, ensure CUDA 12+ is installed. Run nvidia-smi in the Hub terminal to verify driver version.
Poor reconstruction quality (blurry or broken) Common causes: too few overlapping photos, inconsistent lighting, very shiny or transparent surfaces. Add more photos covering problem angles. Avoid reflective surfaces.
Viewer renders slowly Gaussian Splatting is GPU-intensive in the viewer too. Reduce splat count via Project → Settings → Viewer Quality. On lower-end hardware, use the Mesh Export mode for a lighter representation.
Splat not appearing in Spatial Time Machine Confirm Digital Memory integration is enabled in Photogram → Settings → Integrations. Splats appear in the timeline after the sync completes (up to a few minutes after training finishes).