Nodes/ComfyUI-TurntableGSViewer/Gaussian Merge to Target
ComfyUI Node

Gaussian Merge to Target

A ComfyUI node in viewer with 6 inputs and 1 output.

By PozzettiAndrea·Created 3 months ago·Updated 2 months ago· 6
Gaussian Merge to Target
    • ply_path
    ply_path
    target_count200000
    output_filenamemerged
    opacity_threshold0.10
    k16
    backendtorch_gpu
    Categoryviewer

    Inputs (6)

    NameTypeDefaultDescription
    ply_pathSTRINGPath to a Gaussian Splatting PLY file
    target_countINT2000001000–10000000Approximate output Gaussian count. NanoGS works in ratio terms - target / input_count is passed as `--ratio`. If target >= input count, the input is passed through unchanged.
    output_filenameSTRINGmergedBasename for the output PLY (no extension).
    opacity_thresholdFLOAT0.100–1Splats with opacity below min(threshold, median(opacity)) are pruned before merging.
    kINT164–64k nearest neighbors per splat that the merge graph considers as candidate pairs. Higher k -> slightly better quality (more options for the greedy selector) but more RAM + slower (O(N*k) edges). 16 is the upstream sweet spot. 8 = fast draft. 24-32 helps on specular-heavy scenes; 32+ rarely pays off.
    backendCOMBOtorch_gputorch_gpu - MPMM on GPU via PyTorch (CUDA / ROCm / MPS, falls back to CPU torch). ~5-10x faster than nanogs_cpu. nanogs_cpu - original NanoGS NumPy/SciPy backend.

    Outputs (1)

    NameTypeDescription
    ply_pathSTRING