Running Lumen

The Lumen container image: GPU and driver requirements, the data volume, fetching the encoder weights, and serving the web console.

Lumen ships as a single container image that carries the server, the reconstruction pipeline and the web console. The only host requirements are a container engine (Docker or Podman), an NVIDIA GPU and the NVIDIA container toolkit.

Requirements

GPU

An NVIDIA GPU with hardware video decode (NVDEC). Reconstruction quality does not depend on the card, but training time does — a modern high-memory card trains a full scene in minutes.

Driver

A CUDA 13 capable NVIDIA driver on the host, with the NVIDIA container toolkit installed. The image requests the driver’s video capability, which is what makes hardware decode available inside the container.

Storage

One volume, mounted at /opt/lumen/data. Stores, staged clips and model weights all live under it; sizing depends on the corpus — budget on the order of a gigabyte per reconstructed scene.

First run

Verify the GPU stack, fetch the encoder weights (~600 MB, one time, into the data volume), then serve:

docker run --rm --gpus all metaloom/lumen gpu-doctor
docker run --rm --gpus all -v lumen-data:/opt/lumen/data metaloom/lumen fetch-models
docker run -d --gpus all -p 8080:8080 -v lumen-data:/opt/lumen/data --name lumen metaloom/lumen

gpu-doctor audits every layer the pipeline depends on — driver, kernel compilation, hardware decode, inter-process sharing — and exits non-zero naming the first thing that failed, which is considerably more useful than a stack trace from the middle of an ingest.

The console is then reachable on port 8080 of the host. Without the encoder weights the corpus still browses; the search box itself explains what is missing rather than failing silently.

Exposure and authentication

The server carries no authentication. Outside a container it therefore binds to the local machine only; the container image overrides this so the published port works, and the server prints a warning saying so at startup. Publish the port to localhost, a private network or behind your own authenticating reverse proxy — never directly to the internet.

Day-to-day operation

Ingest through the console. The Ingest panel accepts clips and queues them behind the single GPU worker; progress streams per stage.
Everything is also headless. Every operation the console offers — reconstruction, import, search, rendering, corpus audit — exists as a command inside the image, so batch work and scripting need no browser session.
The store survives the container. All state lives on the data volume; the container itself is disposable and upgrades are a matter of running a newer image against the same volume.

Looking for something else?

All documentation Feature overview