🔬 WorldMirror 3D (Official)

If VNCCS_WorldMirror3D is the full-featured V1 reconstruction node, _Official is the pared-back version: same model, same outputs, fewer knobs, closer to what Tencent's own release does. The "Official" in the name is doing a specific job - this node sticks closer to the reference pipeline, minus the pack author's hidden-parameter fixes that make the basic pipeline actually hold together. It's the node you reach for when you want the plain, vanilla WorldMirror behavior to compare against, or when you want a leaner node that's less likely to confuse a beginner.

The honest framing

Here's the thing the pack author said out loud in the release thread: the basic WorldMirror V1 pipeline shipped upstream was "a bit broken" - without the extra parameters this pack adds, you'd get "just separate 4 walls with gigantic seam lines." The non-official VNCCS_WorldMirror3D includes those fixes (distortion masking, consensus merging, edge filtering, panorama lock). _Official deliberately does not include most of them. So "Official" here means faithful to the reference, not better. If your output has seam lines or ghosting, the answer is usually to switch to the plain VNCCS_WorldMirror3D node with its mask and consensus settings, not to tweak this one.

What it gives you

Required: model (a WORLDMIRROR_MODEL from the V1 loader) and images. The optional set is small and sensible:

• target_size - inference resolution, default 518. The standard sweet spot.
• use_gsplat - splat output on; disable for point cloud only.
• use_direct_points - use the model's pointmap head (pts3d) instead of depth reprojection. The tooltip's honest about it: "may improve multi-view overlap alignment." Cheap to try.
• confidence_percentile - drops the bottom X% low-confidence points. Default 10 is a fine starting place.
• filter_edges + edge_depth_threshold - strips points at depth discontinuities.

Outputs are the same six as the main node: ply_data, depth_maps, normal_maps, camera_poses, camera_intrinsics, raw_splats. It plugs into SavePLY and SplatRefiner identically.

Bottom line

Use this when you want the reference behavior and a clean mental model of what the stock pipeline does - great for learning the ropes or A/B-ing against the fixed version. If you want the version that's tuned to actually produce seamless rooms, you want VNCCS_WorldMirror3D. Both need the same install: Manager search HY-World 2.0, or clone + pip install -r requirements.txt + python install.py, with the vendored gsplat_maskgaussian CUDA fork being the one dependency you must not replace with upstream gsplat.