VirtualNeRFExplorer

VirtualNeRFExplorer is a minimal viser-based scene explorer for Nerfstudio-compatible models.

v1 Scope

Implemented:

• load any model that Nerfstudio can restore from config.yml
• correct path repair for moved configs by rebasing relative paths during load
• live RGB rerender while moving
• depth and accumulation render modes
• lower resolution during motion, higher resolution after motion settles
• origin frame and world grid
• optional sampled training camera frustums
• click a displayed training camera to snap to that pose
• reset/focus/next-camera navigation controls
• saved camera viewpoints
• image capture of the current rendered frame
• minimal GUI for visibility toggles and render resolution

Repository Layout

VirtualNeRFExplorer/
├── pyproject.toml
├── README.md
└── src/
    └── virtual_nerf_explorer/
        ├── __init__.py
        ├── __main__.py
        ├── app.py
        ├── camera.py
        ├── config.py
        ├── loader.py
        ├── render.py
        ├── session.py
        ├── state.py
        └── viewer/
            ├── __init__.py
            ├── explorer.py
            ├── gui.py
            └── scene.py

Architecture

The implementation follows Nerfstudio viewer patterns, but strips them down to only what is required for interactive scene exploration.

• loader.py
  • loads the Nerfstudio config and pipeline through eval_setup(...)
  • repairs relative data and output_dir paths for moved configs
• camera.py
  • converts between viser client camera state and Nerfstudio render camera state
• render.py
  • owns a per-client render worker thread
  • renders low resolution during motion and high resolution after motion stops
• viewer/scene.py
  • builds the world grid, origin frame, and training camera frustums
• viewer/gui.py
  • builds the minimal control panel and status block
• viewer/explorer.py
  • wires client lifecycle, render workers, camera snapping, and GUI callbacks

Install

cd /workspace/Desktop/Repos/VirtualNeRFExplorer
pip install -e .

Run

cd /workspace/Desktop/Repos/VirtualNeRFExplorer
virtual-nerf-explorer \
  --load-config /path/to/config.yml \
  --host 0.0.0.0 \
  --port 8080

Example with your current dataset:

cd /workspace/Desktop/Repos/VirtualNeRFExplorer
virtual-nerf-explorer \
  --load-config /workspace/Desktop/DATASETS/CHERRY_DATASET_2_UOH_O/CHERRY_DATASET_2_UOH/nerfacto/2026-04-16_162613/config.yml \
  --host 0.0.0.0 \
  --port 8080

Main Flags

• --load-config
  • required Nerfstudio config file
• --show-training-cameras / --no-show-training-cameras
  • show or hide sampled training frustums on startup
• --show-world-axes / --no-show-world-axes
  • show or hide origin frame and grid on startup
• --max-display-cameras
  • cap the number of frustums drawn in the viewer
• --min-orbit-distance
  • enforce a minimum camera-to-look-at distance to avoid the orbit controls becoming sluggish near the scene center
• --static-max-res
  • maximum resolution after the camera stops moving
• --moving-max-res
  • maximum resolution while the camera is moving
• --depth-quantile
  • cumulative weight threshold used by the depth view; 0.5 matches Nerfstudio median depth

Notes

• Scene capture is available from the GUI via Capture.
• Capture downloads the latest rendered NeRF frame already shown in the explorer instead of triggering a separate browser-side scene render.
• The Name field lets you choose the download base name safely; invalid filename characters are sanitized automatically.
• Tensors -> Export downloads the current render package as .npz or .pt, including image-space ray tensors plus model outputs such as RGB, depth, accumulation, density, and related arrays when available.
• Render switches between rgb, depth, and accumulation outputs using the current model outputs.
• Depth cumsum adjusts the cumulative-weight split used by the depth view. 0.5 is the standard median depth; lower values bias toward nearer samples and higher values bias deeper into the volume.
• Reset, Focus, Next cam, and Snap provide basic scene-navigation controls without patching viser.
• Saved views are stored in memory for the current explorer session and can be loaded or deleted from the Views folder.
• The current installed viser API does not expose a backend keyboard shortcut hook, so capture is implemented as a button rather than a true key binding.
• Rendering is serialized through a lock to avoid concurrent model access from multiple clients.
• Relative paths in moved Nerfstudio configs are repaired automatically using the config location and inferred run root.
• The package sets TORCHINDUCTOR_COMPILE_THREADS=1 by default before Nerfstudio imports so the viewer does not spawn a large compile-worker process pool and Ctrl+C remains usable.
• This project is intentionally simpler than the full Nerfstudio viewer. The current product is a scene explorer, not a training UI.