WorldGen Build GS Data From WorldMirror

WorldMirror's splats are feed-forward - fast, but rough. The proper fix is to take your reconstruction's images and cameras and train a real Gaussian splatting scene, the way the 3DGS literature does it. VNCCS_WorldGenBuildGSDataFromWorldMirror is the bridge: it takes your ply_data, images, and cameras, and writes out a gs_data folder in the standard native-3DGS trainer format - images, aligned point cloud, camera poses and intrinsics - that the pack's trainer (and the vendored gsplat fork) can consume.

Required inputs: ply_data, images, camera_poses, camera_intrinsics - all straight off a reconstruction node. Optional: a workspace_dir from WorldGenExportBankFromPLY (omit it and a default workspace is created), plus depth_maps, normal_maps, and raw_splats if you have them.

The inputs that matter

• out_name (default gs_data) - the subfolder inside the workspace where the trainer dataset lands. Connect this node's gs_data_dir output to whatever trainer node follows.
• camera_bundle - the honest knob. Default is model_predicted_when_input_zero: when your connected cameras have zero translation (e.g. you only fed the model a single view or rotation-only panoramas), it swaps in WorldMirror's predicted cameras/geometry instead. The tooltip says why in one line: "Native 3DGS needs camera baseline; rotation-only cameras collapse depth." If your source cameras genuinely have translation, set it to connected_inputs.
• points_max (default 3M) - cap on the point cloud written for the trainer. Lower = smaller/faster, less detail.
• write_normals (default true) - writes normal data if available; some trainers use it.
• deep_logging - verbose diagnostics to the console.

Outputs: gs_data_dir (a string path - wire it onward) and info.

Where this fits

This is the "do it properly" path: WorldMirror gives you a rough splat, this node packages the raw material for a native 3DGS refinement that actually converges. The KB's lesson applies directly - a trained 3DGS scene is the right deliverable for an explorable scene, and forcing it to a mesh throws the quality away. Expect the training step itself to be slow and VRAM-hungry (it's a real optimization loop); that's the price of clean geometry, and the fork-only rasterization arguments (distloss, gauss_masks) are exactly why this pack insists its vendored gsplat be the only one installed.

Install

Pack standard - Manager → HY-World 2.0, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AHEKOT/ComfyUI_HYWorld2
cd ComfyUI_HYWorld2
pip install -r requirements.txt
python install.py

The install.py step (build the vendored gsplat_maskgaussian CUDA fork) is non-negotiable here - this node's whole downstream is gsplat.