Gaussians From Point Cloud

Most of the 3D stuff you generate in ComfyUI doesn't come out as gaussian splats. It comes out as a point cloud - a PLY with just x/y/z and maybe r/g/b. Fine for a point viewer, useless for a gaussian-splat pipeline, because 3DGS needs a whole set of per-point fields: rot_*, scale_*, opacity, and f_dc_* color. Feed a bare point cloud into a splat trainer or a viewer without those fields and you get either a rejection or degenerate defaults - the classic failure being every gaussian collapsed to a 1 mm sphere, which then takes forever to train out of.

GaussiansFromPointCloud is the bridge. It takes any PLY - plain point cloud or an existing 3DGS PLY - and emits a properly-initialized gaussian PLY that downstream trainers and viewers actually like. It's part of the GaussianPack family (repo ComfyUI-TurntableGSViewer, same author as the HyWorld2 and TRELLIS.2 wrappers) and lives in the "GaussianPack" menu category.

How it works

The key is that scales aren't just guessed - they're derived the same way upstream 3DGS does its SfM initialization. The node runs a k-nearest-neighbors lookup (via sklearn's KDTree) and sets each gaussian's scale to log(mean distance to its k nearest neighbors × init_scale). Dense regions get small gaussians, sparse regions get big ones. That's exactly the initialization a real trainer would converge from, so a later training pass starts from a sane place instead of a soup of 1 mm spheres.

The rest of the fields follow sensibly:

• colors from the PLY's r/g/b flow into f_dc_* via the standard inverse transform
• rotations start as identity quaternions (training deforms them)
• opacity starts at logit(init_opacity), default 0.1 - which matches the upstream SfM init

Inputs that matter

• ply_path - input PLY. Plain point clouds get the full treatment; existing 3DGS PLYs have all fields read and missing ones defaulted. Note: knn-derived scales override whatever scales the input already had.
• init_scale (default 1.0) - multiplier on the knn distance. Below 1 gives sharper but possibly under-covered gaussians; above 1 gives smoother but blurrier ones. 1.0 is the honest default.
• init_opacity (default 0.1) - the sigmoid-space initial alpha. The node writes logit(init_opacity) to the opacity field.
• knn_k (default 3) - neighbors used for scale estimation, matching upstream. Larger k = smoother scale field.
• subsample_max (default 0) - optional random cap on output count. 0 means write every point. If a downstream node OOMs on a giant cloud, cap it at something like 2,000,000; the seed is deterministic from the input count.
• output_filename and output_dir - filename, and where to write (leave the dir blank for ComfyUI's output folder).

The single output, ply_path, is the absolute path of the new PLY - ready for a viewer, merge, export, or a trainer.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-TurntableGSViewer.git
cd ComfyUI-TurntableGSViewer
pip install -r requirements.txt --upgrade
python install.py

Restart ComfyUI. Manager users: search "GaussianPack". Remember the experimental comfy-env dependency pulls in pixi on first install. No weights to download; the pack auto-copies sample PLYs into input/.

When you'll use it

Any time your upstream is point-cloud-flavored - a depth-to-pointcloud node, a colmap export, a panorama-stitched scene - and your downstream expects real gaussians. It's also a handy sanity fix for an existing 3DGS PLY whose scales are garbage. One thing to watch: a 7M-point cloud takes a few seconds and a couple hundred MB of RAM for the knn pass, so don't be alarmed if it's not instant.