Command Line Options - MrNeRF/LichtFeld-Studio GitHub Wiki

  TRAINING PATHS:
    -d[data_path], --data-path=[data_path]
                                      Path to training data
    -o[output_path], --output-path=[output_path]
                                      Path to output
    --config=[config_file]            LichtFeldStudio config file (json)
    --init=[path]                     Initialize from splat file (.ply, .sog, .spz, .resume)
    --import-cameras=[path]           Import COLMAP cameras from sparse folder (no images required)

  TRAINING PARAMETERS:
    -i[iterations], --iter=[iterations]
                                      Number of iterations
    --strategy=[strategy]             Optimization strategy: mcmc, adc, igs+
    --sh-degree=[sh_degree]           Max SH degree [0-3]
    --sh-degree-interval=[sh_degree_interval]
                                      SH degree interval
    --max-cap=[max_cap]               Max Gaussians for MCMC or igs+
    --min-opacity=[min_opacity]       Minimum opacity threshold
    --steps-scaler=[steps_scaler]     Scale training steps by factor
    --tile-mode=[tile_mode]           Tile mode for memory-efficient training: 1=1 tile, 2=2 tiles, 4=4 tiles (default: 1)

  INITIALIZATION:
    --random                          Use random initialization instead of SfM
    --init-num-pts=[init_num_pts]     Number of random initialization points
    --init-extent=[init_extent]       Extent of random initialization

  DATASET OPTIONS:
    --images=[images]                 Images folder name
    --test-every=[test_every]         Use every Nth image as test
    -r[resize_factor], --resize_factor=[resize_factor]
                                      Resize resolution by factor: auto, 1, 2, 4, 8 (default: auto)
    --max-width=[max_width]           Max width of images in px (default: 3840)
    --no-cpu-cache                    Disable CPU memory caching (default: enabled)
    --no-fs-cache                     Disable filesystem caching (default: enabled)
    --undistort                       Undistort images on-the-fly before training

  MASK OPTIONS:
    --mask-mode=[mask_mode]           Mask mode: none, segment, ignore, alpha_consistent (default: none)
    --invert-masks                    Invert mask values (swap object/background)
    --no-alpha-as-mask                Disable automatic alpha-as-mask for RGBA images

  SPARSITY OPTIMIZATION:
    --enable-sparsity                 Enable sparsity optimization
    --sparsify-steps=[sparsify_steps] Number of steps for sparsification (default: 15000)
    --init-rho=[init_rho]             Initial ADMM penalty parameter (default: 0.0005)
    --prune-ratio=[prune_ratio]       Final pruning ratio for sparsity (default: 0.6)

  RENDERING OPTIONS:
    --enable-mip                      Enable mip filter (anti-aliasing)
    --bilateral-grid                  Enable bilateral grid filtering
    --ppisp                           Enable PPISP for per-camera appearance modeling
    --ppisp-controller                Enable PPISP controller for novel views
    --bg-modulation                   Enable sinusoidal background modulation
    --gut                             Enable GUT mode

  OUTPUT OPTIONS:
    --eval                            Enable evaluation during training
    --save-eval-images                Save evaluation comparison images (GT vs rendered)
    --save-depth                      [TODO] Save depth maps during training (not yet implemented)
    --timelapse-images=[timelapse_images...]
                                      Image filenames to render timelapse images for
    --timelapse-every=[timelapse_every]
                                      Render timelapse image every N iterations (default: 50)

  UI OPTIONS:
    --headless                        Disable visualization during training
    --train                           Start training immediately on startup
    --no-interop                      Disable CUDA-GL interop (use CPU fallback for display)
    --debug-python                    Start debugpy listener on port 5678 for plugin debugging
    --debug-python-port=[port]        Port for debugpy listener (default: 5678)

  LOGGING:
    --log-level=[level]               Log level: trace, debug, info, perf, warn, error, critical, off (default: info)
    --verbose                         Verbose output (equivalent to --log-level debug)
    -q, --quiet                       Suppress non-error output (equivalent to --log-level error)
    --log-file=[file]                 Optional log file path
    --log-filter=[pattern]            Filter log messages (glob: *foo*, regex: \\d+)

  EXTENSIONS:
    --python-script=[path...]         Python script(s) for custom training callbacks

SUBCOMMANDS:
convert -- Convert between .ply, .sog, .spz, .html
plugin -- Manage plugins (create, check, list)
Run '<subcommand> --help' for details.
EXAMPLES:
lichtfeld-studio -d ./data -o ./output
lichtfeld-studio --resume checkpoint.resume
lichtfeld-studio -v model.ply
lichtfeld-studio convert in.ply out.spz
lichtfeld-studio plugin create my_plugin
ENVIRONMENT:
LOG_LEVEL -- Set log level (trace/debug/info/perf/warn/error)