Lichtfeld Studio

LichtFeld Studio Skill

Control LichtFeld Studio — a native C++23/CUDA workstation for 3D Gaussian Splatting — via its built-in MCP HTTP server on port 45677.

When to Use

• Training 3D Gaussian Splat models from COLMAP datasets
• Rendering / batch multi-view rendering from trained gaussian scenes
• Editing gaussian scenes (selection, deletion, transformation)
• Exporting models (PLY, SOG, SPZ, USD, HTML) or converting between formats
• Automated quality assessment or LLM-guided scene cleanup (floater removal)
• Converting a video into a COLMAP dataset (SplatReady) for training

When Not To Use

• General 3D modeling (meshes, curves) — use the blender skill
• AI image generation from text — use the comfyui skill
• 2D image processing — use the imagemagick skill
• Geospatial 3D — use the qgis skill

Gate

agentbox.toml's [skills.spatial_and_3d] sets gaussian_splatting = false by default. Unlike spark-scene, which discloses its own gate landscape up front, this means the binary and bridge below are not started automatically: build LichtFeld Studio and launch it (or --headless) yourself before the MCP tools below can reach it. Loading this skill still requires neither an image rebuild nor a new MCP daemon — only the manual start.

Architecture

LichtFeld Studio has a built-in MCP server speaking JSON-RPC 2.0 over HTTP POST at http://127.0.0.1:45677/mcp. A stdio-to-HTTP bridge script auto-launches the app.

Claude Code → stdio bridge → HTTP POST → LichtFeld MCP Server (port 45677)

Quick Path

Two runnable helpers ship with this skill — prefer them over hand-rolling curl:

# MCP control — ping / discover / call tools / read resources
tools/lfs-mcp.sh ping
tools/lfs-mcp.sh list                      # live tool discovery (70+ tools)
tools/lfs-mcp.sh call training.get_state
tools/lfs-mcp.sh call render.capture '{"width":1920,"height":1080}'
tools/lfs-mcp.sh read lichtfeld://training/state

# End-to-end: video → COLMAP → trained splat
tools/video2splat.sh input.mp4 ./out [fps] [max_iter] [strategy]

Setup — MCP server config (recommended)

Add to Claude settings to expose all built-in tools as MCP tools:

{
  "mcpServers": {
    "lichtfeld": {
      "command": "python3",
      "args": ["/home/devuser/workspace/gaussians/LichtFeld-Studio/scripts/lichtfeld_mcp_bridge.py"],
      "env": {
        "LICHTFELD_EXECUTABLE": "/home/devuser/workspace/gaussians/LichtFeld-Studio/build/LichtFeld-Studio"
      }
    }
  }
}

Alternatives: talk to a running app directly over HTTP, run --headless (no display), or LichtFeld-Studio convert in.ply out.spz for GPU-free format conversion. Full forms in references/workflows.md.

References

• references/tool-catalog.md — the 70+ MCP tools by category (training, camera, render, selection, scene graph, export, history, crop/ellipsoid, python editor, events, low-level gaussians, plugins) + read-only resource URIs.
• references/workflows.md — raw JSON-RPC HTTP forms, worked workflow examples (train, multi-format export, LLM cleanup, batch render), full CLI reference, and environment variables.
• references/splatready.md — the SplatReady video→COLMAP plugin: pipeline stages, per-stage CLI, output layout, and dependencies.

Troubleshooting

• App won't start: run LichtFeld-Studio --warmup to verify CUDA/PTX compilation.
• MCP not responding: tools/lfs-mcp.sh ping (or curl -s http://127.0.0.1:45677/mcp -d '{"jsonrpc":"2.0","id":0,"method":"ping"}').
• Headless no MCP: the headless path currently does not start the MCP server (GUI-only). Use GUI mode or apply the headless MCP patch.
• Bridge log: check ~/.codex/log/lichtfeld-mcp-bridge.log.