Custom nodes for VOID: Video Object and Interaction Deletion by Netflix Research. VOID removes objects from videos along with all physical interactions they cause (falling objects, displaced items, etc.), not just shadows and reflections.

Built on CogVideoX-Fun-V1.5-5b-InP and fine-tuned with interaction-aware quadmask conditioning.

Paper: arxiv 2604.02296 · GitHub: Netflix/void-model · Model weights: Apache-2.0

Example

| Input | Output | |-------|--------| | Input | Output |

Coffee cup removed from a wooden table using Pass 1 at 672×384, 1 frame, 50 steps, CFG 6.0.

Downloads

| File | Source | Destination | |------|--------|-------------| | void_pass1.safetensors | netflix/void-model | ComfyUI/models/diffusion_models/ | | void_pass2.safetensors (optional) | netflix/void-model | ComfyUI/models/diffusion_models/ | | CogVideoX VAE (diffusion_pytorch_model.safetensors) | alibaba-pai/CogVideoX-Fun-V1.5-5b-InP → vae/ subfolder | ComfyUI/models/vae/ | | T5-XXL text encoder (e.g. t5xxl_fp16.safetensors) | ComfyUI model manager | ComfyUI/models/clip/ |

Pass 1 is sufficient for most videos. Pass 2 adds optical-flow warped noise initialization for improved temporal consistency on longer clips.

The original THUDM/CogVideoX-5b VAE (cogvideo5bvae.safetensors) also works — it shares the same architecture and scaling factor (0.7).

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/your-username/AP_Netflix_VOID

No extra Python packages required beyond a standard ComfyUI installation.

Nodes

VOID Model Loader (AP)

Loads void_pass1.safetensors or void_pass2.safetensors → VOID_MODEL

| Input | Description | |-------|-------------| | model_name | Select from models/diffusion_models/ | | dtype | bfloat16 (recommended) |

VOID VAE Loader (AP)

Loads the CogVideoX 3D VAE → VOID_VAE

| Input | Description | |-------|-------------| | vae_name | Select from models/vae/ | | dtype | bfloat16 (recommended) |

VOID Text Encode (AP)

Encodes a text prompt with a T5-XXL CLIP model → CONDITIONING

Connect a CLIPLoader (T5-XXL) to the clip input.

Prompt tip: Describe the scene after the object is removed — what the clean background looks like. Do not describe the object being removed.

✅ "wooden table, warm lighting" (what should be there after removal)

❌ "A person being removed from the scene."

VOID Quad Mask (AP)

Converts separate mask layers into the 4-level quadmask VOID requires → MASK

A standard binary ComfyUI mask is also accepted and auto-converted (white = remove).

| Input | Type | Description | |-------|------|-------------| | remove_mask | MASK (required) | White = primary object to delete | | affected_mask | MASK (optional) | White = region that physically reacts (e.g. object that falls when person is removed) | | overlap_mask | MASK (optional) | White = pixels that are both primary and affected |

Quadmask pixel values:

| Value | Meaning | |-------|---------| | 0.0 (black) | Primary object — remove | | 63/255 | Overlap of primary + affected | | 127/255 | Affected region (physics interactions) | | 1.0 (white) | Background — keep |

VOID Sampler (AP)

Main inpainting node. Encodes image + quadmask, runs DDIM, decodes output.

| Input | Type | Description | |-------|------|-------------| | void_model | VOID_MODEL | From VOID Model Loader | | vae | VOID_VAE | From VOID VAE Loader | | image | IMAGE | Source frames — resize to 384×672 (H×W) for best results | | mask | MASK | From VOID Quad Mask or any binary mask | | positive | CONDITIONING | Background description after removal | | negative | CONDITIONING | Negative prompt (quality artifacts, watermarks, etc.) | | num_frames | INT | Frames to process. Use 4k+1 values: 1, 5, 9, 13, 17, 49, 97, 197 | | steps | INT | DDIM steps — official default: 50 | | cfg | FLOAT | CFG scale — recommended: 6.0 (matches official VOID pipeline) | | seed | INT | RNG seed | | eta | FLOAT | 0 = deterministic DDIM, 1 = DDPM stochastic | | denoise | FLOAT | Official default: 1.0 (full noise from scratch) |

Outputs: LATENT, IMAGE

VOID Latent → Image (AP)

Re-decodes a VOID LATENT using the 3D VAE without re-running the sampler.

Recommended Workflow

A ready-to-use ComfyUI workflow is included: examples/03.04 Netflix VOID Test.json

Workflow

[CLIPLoader T5-XXL] ──► [VOID Text Encode] ──► CONDITIONING (positive)
                                                                    │
[Load Image]        ──► IMAGE ─────────────────────────────────► [VOID Sampler] ──► IMAGE
[Load Mask]         ──► [VOID Quad Mask] ──► MASK ─────────────►
[VOID Model Loader] ──► VOID_MODEL ────────────────────────────►
[VOID VAE Loader]   ──► VOID_VAE ──────────────────────────────►

Notes

Resolution: 384×672 pixels (H×W) is the training resolution. Other sizes work but quality may degrade.
Frame count: Must be 4k+1 for perfect VAE round-trip: 1, 5, 9, 13, 17, 49, 97, 197. Max: 197 frames.
CFG: Use 6.0 — the official VOID pipeline default. 1.0 disables CFG entirely and results in the model ignoring the positive prompt, producing poor removals.
Mask convention: The VoidQuadMask node outputs black (0) = remove, white (1) = keep. Pass this directly to the sampler. Do not use InvertMask before it.
Text embeddings: Prompts are internally padded to 226 tokens (the training sequence length). No padding node needed.
VAE scaling factor: 0.7 (correct for both THUDM CogVideoX-5b and alibaba-pai CogVideoX-Fun-V1.5).
Text encoder: Must be T5-XXL (4096-dim). CLIP L/G will produce silent conditioning.

License

Custom node code: MIT — see LICENSE VOID model weights: Apache-2.0 — see netflix/void-model