VOID · Video Object Removal

ABOUT THE WORKFLOW

Erase an Object From Video Upload a video and type the name of the object to remove. SAM3 finds and masks it across frames. VOID fills the gap in two passes, keeping motion, lighting, and physical interactions coherent. Shadows, reflections, and objects that depend on the erased subject are handled, not left floating. Two clips are saved: a raw first-pass output and a refined second-pass version.

Model

• VOID (Video Object and Interaction Deletion) by Netflix Research. Released April 2026 under Apache 2.0. Netflix's first open-source AI tool. Built on CogVideoX, fine-tuned for video inpainting with interaction-aware quadmask conditioning. Two-pass pipeline: Pass 1 fills the masked area, Pass 2 refines temporal stability with optical-flow-warped noise. Preferred over Runway 64.8% to 18.4% in early human preference tests.

• SAM3 (Segment Anything 3.1) by Meta. Text-based object detection and segmentation. Finds and masks the named object across frames from a typed description.

• RAFT optical flow model. Estimates frame-to-frame motion for Pass 2's warped noise refinement.

HOW IT WORKS

Step 1. Upload your video The clip you want to clean up. The model processes up to 121 frames at 24 fps by default. Works great with: live action · VFX plates · stock footage · renders

Step 2. Type the object to remove Name the object in the text cue. "Red cape" or "person on the left." SAM3 finds it and masks it across frames.

Step 3. Describe what should fill the gap Write what the scene looks like after the object is gone. "Medieval knight standing with her hands on her hips" fills the area with a plausible continuation.

Step 4. Hit run and download Pass 1 fills the mask with a rough inpainting. Pass 2 refines it for temporal consistency. Both are saved under VOID_Video_Inpainting as RawOutput and Pass2. Ready for: Premiere · DaVinci Resolve · Nuke · After Effects

First time? Leave every setting as-is. The defaults (121 frames · 24 fps · Pass 1: 30 steps · Pass 2: 24 steps · SAM3 threshold 0.5) are the right starting point for almost everyone.

RECOMMENDED SETTINGS

Quick-start guide. Find the goal that matches yours and copy the settings.

• Standard object removal (most people) — 121 frames · 24 fps · 30 + 24 steps · SAM3 threshold 0.5. The right starting point for almost everyone.

• SAM3 is not finding the object — Lower the threshold below 0.5. Lower values catch more but may mask nearby objects. Be more specific in the text cue: "red cape on the person's shoulders" is sharper than "cape."

• SAM3 masks too much — Raise the threshold above 0.5. Higher values require a stronger match before masking.

• The fill flickers or jitters — Pass 2 exists for this. If you turned it off with Skip Pass 2, turn it back on. If it still flickers, check that the FPS constant matches your source clip's frame rate.

• Want to skip Pass 2 for speed — Set the Skip Pass 2 boolean to true. Pass 1 alone is sufficient for most clips. Pass 2 adds temporal stability on longer or more complex scenes.

• Processing fewer or more frames — Change the Frame Count constant. The source loader's frame_load_cap is linked to this value, so they stay in sync.

• The output plays back at the wrong speed — The save nodes are set to 8 fps while the source loads at 24 fps. Set the save node frame rates to match your FPS constant so the output plays at the correct speed.

Prompts: Two text fields matter. The SAM3 text cue names what to erase: "red cape." The positive inpainting prompt describes what should fill the gap: "medieval knight standing with her hands on her hips." The negative prompt lists artifacts to avoid. All three shape the result.

LEARN

📹 Videos

• Intro to Floyo

• ComfyUI 101 Free Course ft. Sebastian Kamph

• Floyo 101 for Team Collaboration

✨ Quick links

• Help docs

• LoRA Training guide

• Discover more workflows

USE CASES

🎬 VFX Object Removal Erase rigging, crew, or unwanted props from a shot while shadows and reflections are removed with the object.

✂️ Content Cleanup Remove logos, signs, or branding from stock footage or archive clips for reuse.

🎭 Interaction-Aware Erasure Remove a person holding something and have the held object fall or vanish realistically rather than float.

📱 Social and Marketing Clean up a clip for a post by removing a distracting element without a manual VFX pass.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Single objects in open scenes with readable backgrounds

• Objects whose shadows and reflections should also vanish

• Clips with steady or slow camera motion

• Descriptive SAM3 text cues that name the object specifically

⚠️ May produce softer results

• Objects that occupy most of the frame

• Fast motion where the object crosses many other elements per frame

• Densely crowded scenes with many overlapping subjects

• Clips longer than 121 frames in a single run

FAQ

What is VOID? VOID stands for Video Object and Interaction Deletion. It is Netflix Research's first open-source AI tool, released April 2026 under Apache 2.0. It removes objects from video while handling the physical interactions they cause, like shadows, reflections, and dependent objects. Built on CogVideoX with interaction-aware quadmask conditioning, it was preferred over Runway 64.8% to 18.4% in early human preference tests.

What is SAM3 and what does it do here? SAM3 is Segment Anything 3.1 by Meta. In this workflow it takes your typed text cue, finds the named object in the video frames, and generates a mask across all frames. That mask tells VOID where to fill. You do not need to draw or paint a mask by hand.

What is the two-pass pipeline? Pass 1 uses the base inpainting model to fill the masked area with plausible content. Pass 2 takes the Pass 1 output, estimates optical flow between frames, and re-runs inference with flow-warped noise to stabilize temporal consistency. Pass 1 alone is sufficient for most clips. Pass 2 fixes flickering and object morphing on harder cases.

Why does the workflow save two videos? The raw output is the Pass 1 result. The Pass2 output is the refined version. Having both lets you compare whether the refinement pass improved or changed the result, and use whichever looks better.

Is VOID free for commercial use? Yes. VOID is released under Apache 2.0, which allows commercial use, modification, and self-hosted deployment with no revenue threshold. SAM3 is released under the Apache 2.0 license as well.

Why is this called one of the heaviest workflows? It loads two VOID transformer checkpoints (Pass 1 and Pass 2), SAM3, a T5-XXL text encoder, a CogVideoX VAE, and an optical flow model. Locally it requires 40GB of VRAM. On Floyo the infrastructure is handled for you.

How to run VOID video inpainting online? You can run VOID video inpainting online through Floyo. No installation, no setup, no 40GB VRAM requirement. Open the workflow in your browser, upload a clip, type the object to remove, and hit run. Free to try.

WHY FLOYO?

Floyo is the only platform with team collaboration for ComfyUI in the browser. You run workflows with no install. You share run history, assets, and models across your team. You pay only when you generate. Floyo supports open-source and closed-source models.

A designer runs an edit and likes the result. A teammate opens that exact run from shared history and keeps going. No file handoffs. No version confusion.

For studios and enterprise teams, Floyo adds private workspaces, pooled resources, and a team usage dashboard. Other ComfyUI cloud tools run for one person at a time. Floyo runs for the whole team, with transparent per-generation costs.

Ready to try it? Upload a video, type the object to remove, and run it. Both passes save automatically.

→ Launch Workflow, Free

Questions? Watch the free course or check the FAQ above.