ComfyUI-VOID
A ComfyUI custom node pack for running VOID (Video Object and Interaction Deletion) inference directly inside ComfyUI workflows. VOID removes objects and their physical interactions — shadows, contact patches, reflections — from video clips using interaction-aware inpainting built on CogVideoX-Fun-V1.5-5b-InP.
Nodes
| Node | Description |
|---|---|
| VOID Load VAE | Loads the CogVideoX-Fun 3D VAE from models/vae/. |
| VOID Load Text Encoder | Loads the T5-XXL text encoder (fp8 supported) from models/clip/. |
| VOID Loader | Loads both VOID transformer checkpoints and assembles the inference pipelines. |
| VOID Inference | Runs Pass 1 base inpainting and optional Pass 2 warped-noise refinement. |
| VOID Quadmask Builder | Compiles a VOID quadmask from SAM2 / SAM3 segmentation masks. |
Requirements
- ComfyUI with the V3 node API (
comfy_api.latest) - diffusers >= 0.33.1 — CogVideoX requires a newer version than ComfyUI ships by default
- Python packages: see
requirements.txt
Install dependencies:
pip install -r requirements.txt
Installation
- Clone this repository into your
ComfyUI/custom_nodes/directory:
cd ComfyUI/custom_nodes
git clone https://github.com/shanef3d/ComfyUI-VOID
- Restart ComfyUI. The VOID nodes will appear in the VOID category.
Model Weights
Download the following files manually and place them in the specified folders:
| Model | Source | File | Save to |
|---|---|---|---|
| CogVideoX VAE (~1.5 GB) | alibaba-pai/CogVideoX-Fun-V1.5-5b-InP | vae/diffusion_pytorch_model.safetensors |
ComfyUI/models/vae/ |
| T5-XXL text encoder (~5 GB, fp8) | Comfy-Org/flux1-dev | t5xxl_fp8_e4m3fn.safetensors |
ComfyUI/models/clip/ |
| VOID Pass 1 transformer (~10 GB) | netflix/void-model | void_pass1.safetensors |
ComfyUI/models/void-model/ |
| VOID Pass 2 transformer (~10 GB) | netflix/void-model | void_pass2.safetensors |
ComfyUI/models/void-model/ |
The T5 tokenizer (~800 KB) is downloaded automatically from HuggingFace on first use and cached locally — no manual step required.
Typical Workflow
VHS Load Video ──────────────────────────────────────────────────┐
↓
SAM3 Video Segmentation → SAM3 Propagate → SAM3 Video Output → VOID Quadmask Builder
↓
VOID Load VAE ──────────────────┐ VOID Inference
VOID Load Text Encoder ──────────┤→ VOID Loader → Pass 1 output ──→ VHS Save Video
└────────────────→ Pass 2 output ──→ VHS Save Video
Mask generation uses ComfyUI-SAM3 — SAM3 Video Segmentation, SAM3 Propagate, and SAM3 Video Output nodes feed directly into VOID Quadmask Builder. SAM2-based mask nodes are also compatible.
Credits
- VOID model: netflix/void-model — paper
- CogVideoX-Fun: alibaba-pai/CogVideoX-Fun-V1.5-5b-InP
- SAM3 ComfyUI nodes: PozzettiAndrea/ComfyUI-SAM3
License
MIT — see LICENSE.
The VOID model weights are released by Netflix under the Apache 2.0 License.