ComfyUI-HY-World2
Open-source integration for ComfyUI enables 3D Gaussian Splatting generation from images, but developer warns it's unstable and limited.
Independent developer AHEKOT has launched ComfyUI-HY-World2, an open-source GitHub project that integrates HY-World's 3D generation capabilities into the node-based ComfyUI interface. This release provides custom nodes for both HY-WorldMirror and the newer HY-World2 models, allowing users to experiment with generating 3D Gaussian Splats from 2D images directly within their local AI workflows. The project represents a significant step toward making advanced 3D content creation more accessible to the open-source AI community.
However, AHEKOT provides crucial caveats: the integration is currently unstable and represents only a partial implementation of HY-World's promised capabilities. The released component focuses on Gaussian Splatting generation and basic 3D model creation, but lacks the sophisticated world-building and character control features demonstrated in promotional videos. Technical challenges include outdated GSplat libraries requiring custom builds for Python 3.12/3.13 with CUDA 13.1 on Windows, high VRAM consumption for decent resolution outputs, and limitations in how well the WorldMirror model assembles final 3D models from multiple camera angles.
The developer is actively seeking community contributions to improve stability and expand functionality, positioning this as an experimental foundation rather than a production-ready tool. This release highlights both the rapid progress in AI-powered 3D generation and the significant technical hurdles that remain in creating consistent, high-quality outputs from 2D inputs.
- Open-source ComfyUI integration enables 3D Gaussian Splat generation from 2D images using HY-World models
- Requires custom Python 3.12/3.13 builds with CUDA 13.1 and significant VRAM for decent resolution outputs
- Developer warns it's unstable and represents only partial implementation of promised HY-World2 capabilities
Why It Matters
Brings experimental 3D generation to local AI workflows, advancing open-source tools for AI-powered content creation despite current limitations.