LightGaussian

Tip

LightGaussian requires a pre-trained 3DGS model as --init_checkpoint.

Prune model

sw-train \
    --source_path /data/MipNeRF-360/bicycle \
    --output_dir /output/lightgs/ \
    --model lightgaussian \
    --optim lightgaussian_prune \
    --init_checkpoint /path/to/point_cloud.ply

Distillate model

sw-train \
    --source_path /data/MipNeRF-360/bicycle \
    --output_dir /output/lightgs/ \
    --model lightgaussian \
    --optim lightgaussian_distill \
    --init_checkpoint /output/lightgs/PRUNE/point_cloud/iteration_5000/point_cloud.ply \
    --teacher_checkpoint /path/to/point_cloud.ply

Evaluate model

sw-eval \
--source_path /data/MipNeRF-360/bicycle \
--model lightgaussian \
--optim lightgaussian_encode \
--eval_mode ENCODE_DECODE \
--sh_degree 2 \
--checkpoint /output/lightgs/DISTILL/checkpoints/ckpt5000.pth

Note encoding stage of lightgaussian requires importance score. Please use checkpoint instead of .ply file. In default setting, distillation stage will automatically save checkpoint file .