🌍 WorldMirror V2 3D Reconstruction Advanced

The Experimental label here isn't marketing drama - it's the actual reason the node exists. The code docstring says it plainly: this WorldMirror V2 variant is kept as a separate node ID so new knobs can evolve without changing saved state on the stable reconstruction node. New features land here first, get tested, and the stable node stays boring. If you want the latest WorldMirror tricks without rebuilding a workflow every week, this is where to look.

It's also the only place in the pack where you get two genuinely useful conveniences: a one-click low_vram_mode and the dense high-res splat upsampler.

What it does differently

Mechanically it's the same feed-forward reconstruction as the rest of the family - images in, depth/normals/point cloud/splats out, all chunked to fit VRAM. Two things set it apart:

low_vram_mode is a single switch that applies the whole low-VRAM profile at once: depth_only head mode, 1 frame per head chunk, 1 GS param chunk, and 8192-token transformer MLP chunks. On a 16GB card with a high target_size, that's the difference between "it runs, slowly" and "it OOMs instantly". On the Advanced node you'd set all four of those by hand; here it's a checkbox.

splat_upsample_mode is the headline experiment: set it to depth_backproject and the node throws away the model's native splats and rebuilds dense, higher-resolution ones from upsampled depth plus high-res RGB. The neat trick is you can run model inference at a small target_size (the tooltip suggests 518) and still get a big splat grid via splat_upsample_size (default 1022) - the expensive part stays cheap, the splat still gets dense.

Inputs worth touching

• model / images / use_gsplat - same as every V2 node. Model comes from VNCCS_LoadWorldMirrorV2Model.
• low_vram_mode - the convenience switch above. Off by default; flip it on before cranking target_size.
• splat_upsample_mode - depth_backproject enables the dense splat path.
• splat_upsample_size - the high-res splat grid size. Bigger = denser, more VRAM at save time.
• splat_upsample_scale_mode - default depth_adaptive grows Gaussian size with depth; footprint_adaptive follows local 3D spacing; hybrid_adaptive uses both. Good starting point, rarely needs touching.
• splat_upsample_max_points - 9M default cap on the upsampled splats, depth-aware so distant geometry isn't the first thing culled (see splat_upsample_cap_far_bias).

Outputs are the full set: ply_data (PLY_DATA → VNCCS_PLYSceneRenderer or VNCCS_SavePLY), depth_maps, normal_maps, camera_poses, camera_intrinsics, raw_splats (VNCCS_SPLAT → VNCCS_SplatRefiner), plus filter_mask and gs_filter_mask for debugging.

Install

Pack-level install, same as its siblings - Manager search "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

install.py builds the vendored gsplat_maskgaussian fork, recast/navmesh, and PyTorch3D from source - CUDA toolkit matching your torch, ninja, and (on Windows) MSVC build tools required. Don't install upstream gsplat; the fork's rasterizer args won't exist there.

Gotchas

• Same display name as Advanced. Both show as "🌍 WorldMirror V2 3D Reconstruction Advanced" in the node menu. The Experimental one is the one with low_vram_mode; the Advanced one swaps that for normalize_camera_poses_to_first. Check the class name before you wire it up.
• If your dense splats have pinholes on visible surfaces, bump splat_upsample_scale a touch (default 0.003) rather than chasing the scale mode.
• Watch the point cap on huge panoramas. At 9M points and a big splat_upsample_size, the save step can get chunky - drop splat_upsample_max_points if file size or the final write is your bottleneck.
• It's experimental for a reason. The splat-upsample path is where the author tests new ideas; if a new knob is flaky, the stable node is the escape hatch, not this one.