AI video researchers
Study counterfactual video editing and whether generative models can preserve causal and physical consistency.
Independent tool overview
VOID at a glance
VOID is an open-source research model from Netflix and INSAIT that removes an object from video and attempts to rewrite the physical interactions caused by that object.
Visit the official VOID site ↗
- Developer
- Netflix and INSAIT researchers
- Primary task
- Physics-aware video object and interaction removal
- Release type
- Open-source research model
- License
- Apache 2.0 code and checkpoints
- Architecture
- CogVideoX-Fun-V1.5-5b-InP foundation
- Local hardware
- 40GB+ VRAM GPU recommended
- Research venue
- ECCV 2026
Overview
What VOID is
VOID, short for Video Object and Interaction Deletion, tackles a harder problem than ordinary video inpainting. When an object is removed, it also tries to regenerate consequences such as another object falling, a collision not occurring, or a trajectory changing.
The workflow combines user-selected points, SAM-based segmentation, vision-language reasoning, interaction-aware quadmasks, and a CogVideoX-based diffusion model. An optional second pass uses warped noise to improve temporal consistency when the first output contains object-morphing artifacts.
This is a research implementation rather than a polished commercial editor. Netflix provides code and two model checkpoints under Apache 2.0, plus a community-hosted browser demo, but local use requires technical setup, a Gemini API key for the supplied mask pipeline, and a GPU with at least 40GB of VRAM.
Use cases
Who VOID is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Technical video teams
Prototype object removal where shadows, contact, collisions, or dependent motion must also be regenerated.
Open-source experimentation
Inspect, adapt, or benchmark a fully released two-pass video-inpainting pipeline and checkpoints.
Specialized post-production R&D
Test difficult removal shots that go beyond filling the pixels directly behind an object.
Capabilities
Core VOID features
Interaction-aware removal
Attempts to remove both the selected object and its causal effects on the rest of the scene.
VLM-guided quadmasks
Uses vision-language reasoning and segmentation to separate the removed object, affected regions, overlaps, and protected background.
Two-pass inference
Runs a base inpainting pass, with an optional warped-noise refinement pass for improved shape and temporal stability.
Released checkpoints
Provides separate Pass 1 and Pass 2 checkpoint files through the official Hugging Face repository.
Editable mask workflow
Includes point selection, automatic mask generation, and a manual quadmask editor for correcting problem frames.
Process
How the VOID workflow works
Step 1
Select the object
Load a source video and click points on the object that should be removed.
Step 2
Generate the quadmask
Run SAM segmentation and the Gemini-assisted reasoning pipeline to identify both the object and interaction-affected regions.
Step 3
Run Pass 1
Provide the source, quadmask, and a prompt describing the clean remaining background to generate the counterfactual video.
Step 4
Inspect the output
Check whether the object, downstream interactions, and unaffected portions of the video remain coherent.
Step 5
Refine if needed
Correct the mask manually or run Pass 2 when morphing and temporal instability remain.
Cost
VOID pricing and free plan
VOID has no subscription price. The official code and checkpoints are released under Apache 2.0, while users supply their own compute and any third-party API usage.
Code and checkpoints
Free
Download and run the official implementation under the Apache 2.0 license.
- Two 11.1GB checkpoints
- Notebook and inference scripts
- No hosted support or service guarantee
Self-hosted operation
Usage-based costs
Pay for the GPU, storage, hosting, and external model APIs used by your deployment.
- 40GB+ VRAM recommended for inference
- Gemini API key used by the supplied mask reasoner
- Engineering and operations costs are separate
Pricing checked . Check current pricing at the source ↗
Assessment
VOID strengths and limitations
Where it stands out
- Addresses causal interactions instead of treating removal as simple hole filling
- Official code, checkpoints, training utilities, and research paper are publicly available
- Apache 2.0 licensing supports experimentation and adaptation
- Manual mask correction and an optional refinement pass provide more control than a single black-box generation
What to consider
- It is a research pipeline, not a streamlined editor for nontechnical users
- Local inference requires a high-memory NVIDIA GPU and substantial setup
- Automatic mask generation depends on multiple components, including SAM and a Gemini API key
- Object morphing and temporal artifacts can still require a second pass or manual mask edits
- Results are counterfactual generations, so physical plausibility is not guaranteed on every real-world clip
Compare
VOID alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Content Creator
Runway Aleph
Choose Runway Aleph for a commercially hosted, prompt-driven workflow for editing existing footage.
Explore Runway Aleph →Content Creator
Runway Creative
Choose Runway Creative for a broader web-based video generation and editing suite with less technical setup.
Explore Runway Creative →Agents
MiniMax Agent
Consider MiniMax's hosted AI products when managed access matters more than running an open research pipeline locally.
Explore MiniMax Agent →Questions
VOID FAQs
What is the VOID AI video model?
VOID is a video object and interaction deletion framework from Netflix and INSAIT researchers. It removes a selected object and attempts to regenerate the scene as if that object and its physical influence had not been present.
Is VOID free?
The official code and model checkpoints are available under Apache 2.0 at no license fee. You still pay for the GPU infrastructure, storage, engineering work, and any third-party API use needed to run it.
Can I use VOID in a web browser?
A community-hosted Hugging Face Space is linked from the official repository. The main release, however, is designed for technical users who can run the model and its mask pipeline themselves.
How is VOID different from ordinary object removal?
Ordinary inpainting usually fills the region behind an object and may remove effects such as shadows. VOID also targets downstream interactions, such as making a supported object fall or preventing a collision after the selected object is removed.
What hardware does VOID require?
The official quick start recommends a GPU with at least 40GB of VRAM, such as an NVIDIA A100. The authors report training on eight A100 80GB GPUs, but that training requirement is separate from inference.
Is VOID ready for production video editing?
It is better viewed as a research starting point. Teams should test output consistency, security, third-party dependencies, compute cost, and failure handling before using it in production workflows.
Bottom line
Our VOID verdict
VOID is a compelling open research option for physics-aware object removal, especially when deleting an object should also change what happens around it. Its value is in experimentation and specialized R&D; creators wanting a fast, supported editor will be better served by a hosted commercial tool.
Visit VOID website ↗Similar Content Creator Tools
View all Content Creator AI tools →
Veo 3.1 Lite
Veo 3.1 Lite - Google's new budget-friendly video generation mode
MAI-Image-2
MAI Image 2 - Microsoft AI's image model with upgraded photorealism and creativity
Phota Studio
Phota Studio - Phota's personalized photo editing and generation model that preserves your identity across edits
Avatar V
Avatar V - HeyGen's new AI avatar model that generates studio-quality, hyperrealistic videos from a 15-second clip
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.