ComfyUI-SegviGen

A ComfyUI implementation of SegviGen, providing precise 3D part segmentation for SegviGen.

Installation

1. Clone this repository to your ComfyUI/custom_nodes/ folder.
2. Install the required base dependencies:

   pip install -r requirements.txt

3. Run the specialized installation script to handle custom CUDA wheels and internal libraries:

   python install.py

Key Features

• Automated Model Downloading: All model components (Trellis base, SegviGen checkpoints, BiRefNet, DinoV3) are automatically downloaded on first use.
• Decoupled Loaders: Individual control over Shape/Tex Encoders and Decoders for optimal VRAM management.
• Memory Management: Built-in support for DMA-based loading (load_torch_file), RAM-safe initialization (init_empty_weights), and proactive cache clearing.
• Granular Pipeline: Modular nodes for preprocessing, conditioning, sampling, and post-processing (VXZ, Latent Slats, Voxel, GLB).

Node Overview

• SegviGen Shape/Tex Loader Nodes: Targeted loaders for individual TRELLIS components.
• SegviGen Checkpoint Loader: Loads the SegviGen flow checkpoints (Full/Interactive).
• SegviGen Image Preprocessor: Automates background removal (BiRefNet) and image preparation.
• SegviGen Image To Cond: Generates conditioning embeddings (DinoV3).
• SegviGen Sampler: Performs the core texture sampling with optional point-based guidance.
• SegviGen Slat To Voxel: Decodes texture latents to voxel representation.

Acknowledgements

Original SegviGen implementation by fenghora. Uses TRELLIS 2.0 by Microsoft.