🌍 WorldMirror V2 3D Reconstruction Advanced

This is the "I want to see every parameter" version of WorldMirror V2. If you just want a point cloud and a splat, the plain VNCCS_WorldMirrorV2_3D node is your friend. This one exists for when you're fighting a specific artifact, feeding it panorama slices, or squeezing a 360° scene onto 16GB of VRAM. Same model, same guts - every dial exposed.

What it actually does

WorldMirror V2 is the reconstruction half of Tencent's HY-World 2.0. Feed it one image - or several views cut from an equirectangular panorama by VNCCS_Equirect360ToViews - and it predicts depth maps, normal maps, camera poses, and a Gaussian splat, then returns a PLY point cloud plus a VNCCS_SPLAT you can render or refine. This is not the "walk around inside a generated world" part of HY-World; the author is blunt that Tencent only shipped the 3D reconstruction piece, and that's what this node wraps. The output is a scene you orbit in a splat viewer - exactly the job gaussian splatting is good at.

There's no diffusion loop here. It's a feed-forward model: per-image heads predict depth, points, normals, and Gaussian parameters in one pass, then the node backprojects depth into world space (using your camera priors or the model's predicted cameras) and assembles the splat. The knob-heavy part is that everything runs in chunks - frames per head, tokens per transformer MLP - so you can trade speed for VRAM.

The inputs that matter

You can safely ignore most of the ~30 widgets here. These are the ones you'll actually touch:

• model - from VNCCS_LoadWorldMirrorV2Model (downloads tencent/HY-World-2.0, roughly 5GB, on first run; pick bf16 for ~2× over fp32, fp8 if you have an Ampere+ card and torchao).
• images - one or more views. Single flat images work best; the author's own caveat is that WorldMirror still assembles multi-camera panoramas imperfectly, so expect seam conflicts there.
• target_size - longest side in pixels, default 952. The author has pushed this to ~1400 on a 16GB card (his own dev rig is a 5070 Ti), but only with the VRAM levers below.
• head_compute_mode - all computes every head; depth+gs skips points/normals to raise the VRAM ceiling; depth_only skips the native Gaussian output entirely. This is your biggest single VRAM lever.
• offload_scheme - model_cpu_offload moves weights to CPU between uses. Slower, but it's how 16GB survives big target sizes.

Two worth knowing further down: filter_splats defaults to off because applying the filter masks to splats carves black holes into panorama coverage, and apply_sky_mask uses V2's own predicted depth mask - no ONNX model needed.

Outputs and where they go

The node emits eight outputs. The ones you'll wire up:

• ply_data (PLY_DATA) - the point cloud; feed it to VNCCS_PLYSceneRenderer to render from arbitrary camera angles, or VNCCS_SavePLY to write it to disk.
• depth_maps and normal_maps (IMAGE) - handy for a quick look or as inputs elsewhere.
• camera_poses / camera_intrinsics (TENSOR) - the cameras that align the splat; other nodes in the pack expect them paired with the reconstruction.
• raw_splats (VNCCS_SPLAT) - pipe into VNCCS_SplatRefiner or the HYWorld2 memory/alignment nodes.
• filter_mask / gs_filter_mask - debugging views of what got filtered.

Install

Same story as every node in this pack. Easiest route:

# via ComfyUI Manager: search "HY-World 2.0" and Install
# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/AHEKOT/ComfyUI_HYWorld2
cd ComfyUI_HYWorld2
pip install -r requirements.txt
python install.py

The catch: install.py compiles a vendored fork of gsplat (gsplat_maskgaussian) plus recast/navmesh and PyTorch3D from source. It needs a CUDA toolkit roughly matching your torch, MSVC build tools on Windows, ninja, and git. Do not install upstream PyPI gsplat - the project requires exactly one gsplat package and the fork's rasterizer args (distloss, gauss_masks) aren't in the upstream release. If the build fails, python scripts/build_gsplat.py retries it and prints a CUDA smoke test so you know it worked.

Where people get burned

• This node has no low_vram_mode toggle. That convenience switch lives on the Experimental node. Here you set head_compute_mode, chunk sizes, and offload yourself - at 1400+ on 16GB you basically must.
• The display name lies. Both this and VNCCS_WorldMirrorV2_3D_Experimental render as "🌍 WorldMirror V2 3D Reconstruction Advanced" in the menu. The Advanced one adds normalize_camera_poses_to_first (official-pipeline camera normalization for prior poses); the Experimental one adds low_vram_mode. When you search for one, make sure you grabbed the right class name.
• Panorama seams are a known weakness. If you're getting mismatched geometry at view boundaries, try dropping camera_conditioning to intrinsics_only or none - the tooltip notes panorama poses are rotation-only, so prior poses can actively cause the seam conflicts.