MesonGS

MesonGS: Post-training Compression of 3D Gaussians via Efficient Attribute Transformation

Tip

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

Compress and finetune

sw-train \
    --source_path /data/MipNeRF-360/bicycle \
    --output_dir /output/mesons/ \
    --model mesongs \
    --optim mesongs \
    --init_checkpoint /output/gs/point_cloud/iteration_30000/point_cloud.ply

Evaluate

sw-eval \
    --source_path /data/MipNeRF-360/bicycle \
    --output_dir /output/mesongs/ \
    --model mesongs \
    --optim mesongs \
    --eval_mode ENCODE_DECODE \
    --checkpoint /output/mesongs/checkpoints/ckpt5000.pth