TensorBoard Export is a LichtFeld Studio plugin that writes full training metrics from the training lifecycle hooks to TensorBoard-compatible event files and CSV.

It exists because LichtFeld Studio's built-in loss plot is intentionally short-lived UI state. This plugin records the training signal at the source, so long runs can be inspected later in TensorBoard or external analysis tools.

• Writes TensorBoard event files via tensorboardX.
• Writes metrics.csv next to each event file run.
• Writes config.json with plugin/schema version, trainer state, dataset params, and optimization params.
• Starts a new unique run automatically on training start.
• Flushes and closes the current run on training end.
• Exposes plugin UI controls for status, run rotation, folder opening, TensorBoard command copying, and TensorBoard URL opening.
• Avoids known empty metrics: train/progress is written only when max iterations are known, and PSNR is not exported because this LichtFeld runtime commonly reports it as an empty zero signal.

LichtFeld export panel:

TensorBoard dashboard:

• How to use Loss curves in LichtFeld Studio

• LichtFeld Studio >=0.4.2
• LichtFeld plugin API >=1,<2
• Python >=3.12
• tensorboardX is used by the plugin for event file writing.
• TensorBoard itself must be installed and started separately.

If you use a local coding agent such as Codex, Claude Code, or another assistant with filesystem access, you can ask it to install the plugin for you:

Install the LichtFeld Studio plugin from https://github.com/Dok11/lichtfeld-tensorboard-export.

Clone it into the LichtFeld plugin directory with the exact folder name tensorboard_export:
- Windows: %USERPROFILE%\.lichtfeld\plugins\tensorboard_export
- Linux/macOS: ~/.lichtfeld/plugins/tensorboard_export

After installation, run the LichtFeld plugin check if LichtFeld-Studio.exe is available. Do not rename the plugin folder. Tell me the installed path and whether validation passed.

The agent should clone the repository into LichtFeld's plugin directory as tensorboard_export, so the installed plugin root contains pyproject.toml and __init__.py.

Clone this repository into LichtFeld's plugin directory with the plugin directory name tensorboard_export:

git clone https://github.com/Dok11/lichtfeld-tensorboard-export.git "$env:USERPROFILE\.lichtfeld\plugins\tensorboard_export"

On Linux/macOS:

git clone https://github.com/Dok11/lichtfeld-tensorboard-export.git ~/.lichtfeld/plugins/tensorboard_export

The plugin directory must contain pyproject.toml and __init__.py at its root.

To update an existing installation:

cd "$env:USERPROFILE\.lichtfeld\plugins\tensorboard_export"
git pull

On Linux/macOS:

cd ~/.lichtfeld/plugins/tensorboard_export
git pull

Then validate it from the directory that contains LichtFeld-Studio.exe:

$env:PYTHONUTF8='1'
.\LichtFeld-Studio.exe plugin check tensorboard_export

PYTHONUTF8=1 works around a Windows plugin validator issue where dependency files inside .venv can be read with the legacy console code page.

After the plugin files are installed, open LichtFeld Studio, go to Tools > Plugin Marketplace, find tensorboard_export, and click Install.

1. Open LichtFeld Studio.
2. Open the plugin manager and load tensorboard_export.
3. In the TensorBoard Export panel, enable Enabled.
4. Start training.
5. The plugin automatically creates a new run directory and starts writing metrics.

Default output directory:

%USERPROFILE%\LichtFeldTensorBoardRuns

Run TensorBoard separately:

py -m pip install tensorboard
tensorboard --logdir "$env:USERPROFILE\LichtFeldTensorBoardRuns" --port 6006

Then open:

http://localhost:6006/?darkMode=true

Open TensorBoard opens the browser URL only. It does not start the TensorBoard server.

• Enabled: enables automatic logging.
• Write CSV: writes metrics.csv next to the TensorBoard event files.
• Every N steps: controls how often values are written to disk.
• Log dir: root directory for TensorBoard runs.
• Run name: optional custom run name. Leave empty for an automatic name.
• Rotate run now: closes the current writer and starts a new unique run directory.
• Flush and close run: flushes event files and CSV, then closes the current run.
• Open run folder: opens the current run directory in Explorer.
• Copy TensorBoard command: copies the TensorBoard command for the configured log root.
• Open TensorBoard: opens http://localhost:6006/?darkMode=true.

Automatic run names use:

YYYYMMDD_HHMMSS_<dataset>_<strategy>_<iterations>iter

Fields are omitted when LichtFeld does not expose them. If a directory already exists, the plugin appends a numeric suffix.

Loss:

• train/loss_raw: instantaneous loss at the logged iteration.
• train/loss_mean_since_last_write: average loss across steps observed since the previous write.
• train/loss_min_since_last_write: minimum loss since the previous write.
• train/loss_max_since_last_write: maximum loss since the previous write.
• train/loss_ema_100: exponential moving average updated every training step and written every Every N steps.
• train/loss_ema_500: slower exponential moving average updated every training step and written every Every N steps.

Training:

• train/num_gaussians
• train/progress

Performance:

• performance/iters_per_second

Trainer:

• trainer/elapsed_seconds
• trainer/eta_seconds

Config:

• config/session text summary

Each run directory contains:

events.out.tfevents...
metrics.csv
config.json

config.json includes:

• schema_version
• plugin_version
• trainer_state
• strategy_type
• max_gaussians
• total_iterations
• optimization_params
• dataset_params

• TensorBoard must be started separately.
• Old TensorBoard event files keep old tags after plugin upgrades. Create a new run for a clean dashboard.
• train/progress appears only when total iterations are available through the runtime.

From the directory that contains LichtFeld-Studio.exe:

$env:PYTHONUTF8='1'
.\LichtFeld-Studio.exe plugin check tensorboard_export

From this plugin directory:

.\.venv\Scripts\python.exe -m py_compile .\__init__.py