Netflix open-sources Void, a video object-removal model that accounts for scene physics

Netflix has released the code for Void, a video editing model that removes from the frame not only an object but also the consequences of its presence. If you remove a person holding a guitar, a conventional editor will often leave the instrument "floating." Void tries to rebuild the scene as if the object had never been there: the guitar falls, the cushion springs back, the collision no longer happens. For post-production, this is an important step from simply painting over pixels toward cause-and-effect editing. The work was presented by researchers from Netflix and INSAIT at Sofia University, and the preprint appeared on arXiv on 2 April 2026.

This is precisely the main problem of video inpainting. Most current systems can fill a hole in the frame and fix surface artifacts such as shadows or reflections, but they break down when the removed object physically interacts with the scene. In the paper and demos, Void's authors show typical cases: a person holds an item, a load presses on a cushion, one object collides with another.

After a conventional removal, absurd traces of the scene's original logic remain. Void targets exactly these scenarios and, according to its authors, preserves coherent scene dynamics better than ProPainter, DiffuEraser, Runway, MiniMax-Remover, ROSE and Gen-Omnimatte. In other words, the model does not merely retouch the background; it tries to answer a harder question: what should happen in the frame next if a key object suddenly disappears.

Technically, Void is built on top of CogVideoX-Fun-V1.5-5b-InP from Alibaba PAI and fine-tuned for video inpainting. The base model is a 3D Transformer with 5 billion parameters. The key idea is to replace the binary "remove/keep" mask with a quadmask with four values: the object itself, the overlap zone, the region of affected interactions, and the unchanged background. This gives the model not just a cut-out region but a more structured description of what should change in the scene after removal. The input also includes a text description of the desired background state, and the default inference resolution in the repository is 384 by 672 pixels with clips of up to 197 frames.

Before generating anything, the system must understand not only the boundaries of the object being removed but also which parts of the scene depend on it. The repository includes a separate pipeline for this: SAM2 segments the object, and Gemini helps reason about interaction zones, after which the mask can be manually adjusted, if needed, in a built-in editor.

The authors also added two-pass inference. The first pass performs the main removal and fills in the scene. The second is there not for "beauty" but to fix a specific problem of video diffusion: the gradual deformation of objects between frames. It uses optical flow and warped noise from the first-pass results to stabilize shape and trajectories over long clips.

The dataset is also interesting. Real paired videos of the type "with the object / without the object, but with correct physics" barely exist, so the team built such data synthetically from HUMOTO and Kubric, recalculating the scene's physics after a person or object is removed. For HUMOTO, they used motion-capture scenes and re-simulation in Blender, while Kubric covered collisions and interactions between objects.

The code and weights are open, the repository is released under Apache 2.0, and there is a demo on Hugging Face. However, for a quick launch in Colab, the developers warn up front that a GPU with 40 GB of video memory or more is required, and training ran on eight 80 GB A100s.

Void's practical significance goes beyond a nice demo. For studios and creators, it is a potential cut in weeks of manual work on complex shots, where the task is not just to remove an object but to rewrite how the scene behaves after it disappears. For researchers, it is another signal that video models are starting to move from generating plausible frames to modeling causality.

There is a flip side, though: the less noticeably such tools alter real footage, the higher the demands on verifying the authenticity of material. So Void is both a powerful VFX tool and a reminder that the line between editing and rewriting events is getting thinner.