--- license: apache-2.0 tags: - comfyui - pixal3d - 3d-generation - moge - rmbg --- # Pixal3D & MoGe Models for ComfyUI This repository contains the pre-downloaded and compiled model weights required to run **Pixal3D-ComfyUI** on a Windows/Linux system (including Docker environments). Using these pre-downloaded files prevents runtime downloads and allows offline or air-gapped deployments. --- ## 📁 Repository Folder Structure / Cấu trúc thư mục To run Pixal3D correctly, place the files in this repository into your ComfyUI models directory as follows: ``` ComfyUI/ └── models/ ├── Pixal3D/ │ ├── TencentARC_Pixal3D/ # Shape & Texture SLat checkpoints │ ├── briaai_RMBG-2.0/ # Background removal model │ ├── camenduru_dinov3-vitl16-pretrain-lvd... # DINOv3 ViT pretrain weights │ └── torch_hub/ # Neighborhood Attention Flow (NAF) upsampler weights │ └── hub/ │ └── checkpoints/ │ └── naf_release.pth │ └── geometry_estimation/ # MoGe Camera & Normal estimation models ├── moge_1_vitl_fp16.safetensors └── moge_2_vitl_normal_fp16.safetensors ``` --- ## 🚀 How to Download and Setup / Hướng dẫn cài đặt ### Option 1: Direct Download (Manual) / Tải thủ công 1. Download the `Pixal3D` and `geometry_estimation` folders from this repository. 2. Move them directly to your ComfyUI `models/` directory: * Copy `Pixal3D/` folder to `ComfyUI/models/Pixal3D/` * Copy `geometry_estimation/` folder to `ComfyUI/models/geometry_estimation/` --- ### Option 2: Using Hugging Face CLI / Sử dụng dòng lệnh If you want to pull these models on a new server or another machine via the terminal, use the following commands: ```bash # Install Hugging Face Hub library pip install huggingface_hub # Download Pixal3D models hf download vuong69/pixal3d-models --local-dir ./models --local-dir-use-symlinks False ``` --- ## ⚙️ Model Loader Settings in ComfyUI To ensure ComfyUI uses these local models and does not try to download anything from the internet at runtime, set the following option in your **Pixal3D Model Loader** node: * `download_if_missing` = `false` (since all weights are already locally present in the correct path).