By Connie · This article contains affiliate links. We may earn a commission at no extra cost to you if you sign up through our links.
Netflix Just Released Its First Public AI Model — and It Removes Objects From Video With Physics Awareness
Netflix open-sourced a physics-aware video AI on April 2, 2026 — the first public model from a major streaming company. It removes objects and their causal effects from video.
April 2, 2026 · 7 min read · By Connie
Netflix released void-model on Hugging Face on April 2, 2026 — a 5-billion-parameter video AI that removes objects from footage and erases their physical effects (shadows, reflections, collisions). Built on CogVideoX-Fun-V1.5-5b. Supports 197 frames at 384×672. Requires 40GB+ VRAM. Available on GitHub. Research-oriented, not production-ready. First public AI model Netflix has ever released.
What Netflix Actually Released
Netflix open-sourced void-model on April 2, 2026 via Hugging Face and GitHub. The model is formally named netflix/void-model and is built on a fine-tuned version of CogVideoX-Fun-V1.5-5b — a 5-billion-parameter video generation model developed by Alibaba's research team.
The model's core function is object removal from video footage. You mark a region using a "quadmask" input — a bounding box or segmentation mask — and void-model removes the object in that region across all frames of the video clip. It supports clips up to 197 frames at 384×672 pixel resolution.
What makes this different from existing video inpainting tools is the physics-awareness. Void-model does not simply erase pixels and fill the gap with background texture. It also removes the causal downstream effects of the erased object — shadows the object cast, reflections it produced in shiny surfaces, and physical interactions it caused with other objects in the scene.
Technical Specifications
Void-model is derived from CogVideoX-Fun-V1.5-5b, a video generation model trained by Alibaba's CogVideoX team and released under Apache 2.0. Netflix fine-tuned this model specifically for the object-removal task using its own internal VFX datasets — which almost certainly include footage from Netflix productions.
| Specification | Value |
|---|---|
| Base model | CogVideoX-Fun-V1.5-5b (Alibaba) |
| Parameters | 5 billion |
| Max resolution | 384×672 pixels |
| Max frames | 197 |
| VRAM required | 40GB+ (A100 or equivalent) |
| Input format | Quadmask (bounding box or segmentation) |
| License | Research (not commercial production) |
| Available on | Hugging Face, GitHub |
| Release date | April 2, 2026 |
Why This Is Significant
Netflix has been one of the largest private consumers of AI technology for over a decade — its recommendation engine, thumbnail optimization system, and content delivery infrastructure are all AI-powered. But the company has never released a public AI model until now.
The decision to open-source void-model marks a shift. Netflix is not releasing this as a product — it explicitly describes void-model as "research-oriented, not production-ready." But the release positions Netflix in the AI research community and likely serves as recruiting signal for ML engineers who want to work with proprietary production VFX data.
For the VFX and post-production industry, the physics-aware removal capability addresses a real bottleneck. Current AI inpainting tools (Adobe Firefly, Runway Inpaint, Topaz Video AI) can fill the pixel region an object occupied, but they do not model the object's physical interactions. VFX artists must manually clean up secondary effects, which can take hours per shot. Void-model automates that second pass.
Who Can Actually Use Void-Model Right Now
The 40GB+ VRAM requirement puts void-model out of reach for consumer hardware. A standard gaming GPU (RTX 4090) has 24GB VRAM — not enough. Running the model requires professional data center GPUs like the NVIDIA A100 (40GB or 80GB variant) or H100.
| User Type | Can Run Void-Model? | How |
|---|---|---|
| VFX studio / production house | Yes | On-premises A100/H100 cluster or cloud GPU |
| Individual ML researcher | Yes (via cloud) | Google Colab A100, Lambda Labs, RunPod |
| Game developer / indie studio | No | RTX 4090 = 24GB, below requirement |
| Consumer / creator | No | Consumer hardware insufficient |
| Enterprise AI team | Yes | AWS p4d, Azure NC A100, GCP A100 VMs |
How Void-Model Compares to Existing Video AI Tools
| Tool | Type | Physics-Aware | Access | Hardware Need |
|---|---|---|---|---|
| Netflix void-model | Object removal | Yes | Open source (research) | 40GB+ VRAM |
| Adobe Firefly Video | Inpainting / generative | No | Adobe subscription | Cloud (browser) |
| Runway Gen-3 Inpaint | Inpainting | No | Runway subscription | Cloud (browser) |
| Topaz Video AI | Enhancement / removal | No | One-time purchase | 4GB+ VRAM |
| DaVinci Resolve Magic Mask | Masking + removal | No | Free / Studio license | GPU varies |
Frequently Asked Questions
- Netflix / Hugging Face:
netflix/void-modelmodel card (April 2, 2026) - GitHub: netflix/void-model repository — inference scripts and technical documentation
- Alibaba CogVideoX-Fun-V1.5-5b — base model documentation
- TechCrunch: "Netflix releases its first open-source AI model" (April 2, 2026)
Related reading: DeepMind AlphaEvolve · AI agent use cases 2026 · How to use AI for video editing
Get the best AI tools tips — weekly
Honest reviews, tutorials, and Happycapy tips. No spam.