Extensions/ComfyUI_HYWorld2
ComfyUI Extension

ComfyUI_HYWorld2

A ComfyUI extension with 43 custom nodes.

By AHEKOT·Created 4 months ago·Updated about a month ago· 74
AHEKOT/ComfyUI_HYWorld2
Nodes43
On cloudLocal install
CategoryVNCCS/HYWorld2/Utils, VNCCS/HYWorld2
Stars74
Updatedabout a month ago

Nodes (43)

HYWorld2 Dataset Frame Pairs

Turning finished world renders into a training dataset

VNCCS/HYWorld2/Utils
HYWorld2 GS Data

Preparing the training data the 3D Gaussian Splat trainer eats

VNCCS/HYWorld2
HYWorld2 Klein World Expansion (experimental)

Expanding your world with a ComfyUI model of your choice (experimental)

VNCCS/HYWorld2
HYWorld2 Memory Alignment

Glueing multiple camera passes into one coherent point cloud

VNCCS/HYWorld2
HYWorld2 Memory Bank

The scene's long-term memory, and how HY-World2 stops worlds drifting apart

VNCCS/HYWorld2
HYWorld2 Prepare WorldMirror Batch

Packing your memory bank into the exact format WorldMirror wants

VNCCS/HYWorld2
HY-World Qwen Pano Encoder

Prepping the canvas and conditioning for Qwen panorama expansion

VNCCS/HY-World/Qwen Pano
HY-World Qwen Pano Generate

The actual panorama generator — Qwen-based 360° outpainting

VNCCS/HY-World/Qwen Pano
HY-World Qwen Pano Prompt

The prompt cookbook for Qwen panoramic expansion

VNCCS/HY-World/Qwen Pano
HY-World Qwen Pano Seam Blend

The last 32 pixels — fixing the wrap seam on your 360

VNCCS/HY-World/Qwen Pano
HYWorld2 QwenVL

The VLM that reads your scene and writes the captions the pipeline feeds on

VNCCS/HYWorld2
HYWorld2 Train 3DGS

The slow road to a beautiful splat — training 3D Gaussian Splatting

VNCCS/HYWorld2
HYWorld2 Trajectories

Where the camera paths through your 3D world get planned

VNCCS/HYWorld2
HYWorld2 Workspace

The anchor every HY-World2 scene hangs from

VNCCS/HYWorld2
HYWorld2 World Expansion

The node that makes a still scene explorable

VNCCS/HYWorld2
🧩 Apply Splat Patch From Edited Views

Surgically editing a gaussian splat by painting over views (experimental)

VNCCS/3D/Experimental
👁️ Background Preview

Where you finally get to look at your gaussian splat

VNCCS/3D
Camera Trajectory Builder

Scripting the camera moves before the video model rolls

VNCCS/Video
📦 Decompose PLY Data

Pulling the cameras and point cloud out of a PLY_DATA bundle

VNCCS/3D
🔄 360° Panorama to Views

Slicing a 360 panorama into the perspective views WorldMirror can digest

VNCCS/3D
Export WorldStereo Single Model

Export a WorldStereo model to a single .safetensors file

VNCCS/Video
Install HYWorld2 Dependencies

One node that installs the pack's heavy dependencies for you

VNCCS/Install
Load PLY File

Bring a saved .ply back into the graph with its cameras

VNCCS/3D
🌐 Load WorldMirror Model

The loader that downloads Tencent's WorldMirror for you

VNCCS/3D
🌍 Load WorldMirror V2 Model

Loading the HY-World 2.0 / WorldMirror V2 checkpoint

VNCCS/3D
Load WorldStereoLight Model

The int4 WorldStereo loader for smaller GPUs

VNCCS/Video
Load WorldStereo Model

Loading the WorldStereo pipeline (and its MoGe depth helper)

VNCCS/Video
VNCCS Panorama Mapper

Unwrapping a panorama into flat room walls

VNCCS/3D
🎥 PLY Scene Renderer

🎥 Render your saved splat from any camera angle

VNCCS/3D
💾 Save PLY File

💾 Getting your reconstruction onto disk as a .ply

VNCCS/3D
🧩 Assemble Splat Patches

🧩 The node that stitches your splat repairs back together

VNCCS/3D/Experimental
🧩 Splat Patch POC

🧩 Render a splat, find the holes, prep them for repair

VNCCS/3D/Experimental
🧩 Splat Patch POC 2

🧩 Splat Patch POC 2 — the triage node that scans the whole scene

VNCCS/3D/Experimental
🧩 Splat Patch POC v2

🧩 Splat Patch POC v2 — same job, revised knobs

VNCCS/3D/Experimental
🔥 Splat Refiner (Backprop)

🔥 The backprop pass that cleans up your raw splats

VNCCS/3D
WorldGen Build GS Data From WorldMirror

Assemble a native 3DGS training dataset from your reconstruction

VNCCS/WorldGen
WorldGen Prepare Workspace

Turn a reconstruction into a WorldGen workspace

VNCCS/WorldGen
🏔️ WorldMirror 3D Reconstruction

Turn a single photo into an explorable 3D room with WorldMirror

VNCCS/3D
🔬 WorldMirror 3D (Official)

WorldMirror 3D, the trimmed 'official' variant

VNCCS/3D
🌍 WorldMirror V2 3D Reconstruction

The V2 reconstruction node with splat upsampling built in

VNCCS/3D
🌍 WorldMirror V2 3D Reconstruction Advanced

Single image and panoramas to point cloud, depth, and Gaussian splats

VNCCS/3D/Advanced
🌍 WorldMirror V2 3D Reconstruction Advanced

One-click low-VRAM mode and dense high-res splats

VNCCS/3D/Experimental
WorldStereo Generate

Camera-guided video generation with WorldStereo

VNCCS/Video
Readme

ComfyUI HY-World 2.0 - WorldMirror 3D

ComfyUI custom nodes for 3D scene reconstruction from a single image, panorama, image set, or video using HY-World 2.0 / WorldMirror.

Nodes

Category: VNCCS/3D

| Node | Description | |------|-------------| | VNCCS_LoadWorldMirrorModel | Download and load WorldMirror V1 model | | VNCCS_WorldMirror3D | V1 inference: PLY point cloud, depth, normals, Gaussian splat | | VNCCS_LoadWorldMirrorV2Model | Download and load WorldMirror V2 model | | VNCCS_WorldMirrorV2_3D | V2 inference: PLY point cloud, depth, normals, Gaussian splat | | VNCCS_WorldMirrorV2_3D_Clean | V2 reconstruction defaults for cleaner dense splats | | VNCCS_PLYSceneRenderer | Render PLY scenes from arbitrary camera angles | | VNCCS_SplatRefiner | Refine Gaussian splat data | | VNCCS_DecomposePLYData | Extract XYZ / RGB / normals / opacity tensors from PLY | | VNCCS_SavePLY | Save PLY files to disk | | VNCCS_Equirect360ToViews | Extract perspective views from equirectangular panoramas | | VNCCS_PanoramaMapper | Map panoramas to wall / floor / ceiling projections |

Installation

Via ComfyUI Manager

Search for HY-World 2.0 and click Install. requirements.txt and install.py run automatically.

install.py installs the HY-World vendored gsplat_maskgaussian fork from:

hyworld2/worldgen/third_party/gsplat_maskgaussian

Do not install upstream/PyPI gsplat separately. HY-World worldgen and the native scene trainer require fork-only rasterization arguments such as distloss and gauss_masks; the project must have only one installed Python package named gsplat.

Manual

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

HY-World gsplat_maskgaussian

gsplat_maskgaussian is a CUDA extension and is built from the vendored HY-World source. It replaces upstream gsplat under the same package name.

Required build tools:

  • CUDA Toolkit matching the PyTorch CUDA major/minor version as closely as possible.
  • MSVC C++ Build Tools on Windows with the Desktop development with C++ workload.
  • ninja, listed in requirements.txt.
  • Git, if the vendored GLM headers need to be restored.

Manual rebuild:

# Windows embedded ComfyUI Python
scripts\pipinstall.bat

# Any Python environment
python scripts/build_gsplat.py

Successful installation is verified by a CUDA smoke test that calls gsplat.rendering.rasterization with distloss=True and gauss_masks.

Workflows

Example workflows are in the workflows/ directory:

  • World-single-image.json - single image to 3D scene
  • World-Mirror-panorama.json - equirectangular panorama to 3D scene

Requirements

  • Python 3.10+
  • PyTorch with CUDA
  • NVIDIA GPU for gsplat rendering/training
  • Flash Attention for HY-World model inference where supported:
pip install flash-attn --no-build-isolation