Netflix's VOID AI Erases Actors—and Their Physics

Here's the thing about most AI video erasers: they're really good at making things disappear, but they're absolutely terrible at understanding why those things were there in the first place.

Remove a bowling ball mid-strike, and standard AI tools will happily show you pins falling over for absolutely no reason. Delete a person making a smoothie, and the blender keeps spinning like it's possessed. The object vanishes, but its ghost lingers—not in some spooky supernatural way, but in all the physical interactions that suddenly make no sense.

Netflix just released an open-source framework called VOID (Video Object and Interaction Deletion) that actually tries to solve this problem. And judging by the tests run by the team at Better Stack, it's... complicated.

The Ghost Interaction Problem

Most video inpainting models are essentially supercharged content-aware fill. They analyze the pixels around a masked area and make educated guesses about what should fill the gap. This works fine for static objects—a watermark, a person standing still in the background. But the moment there's physical interaction, the whole illusion collapses.

As the Better Stack team explains it: "Standard AI erasers are basically content-aware fill on steroids. They look at the pixels around the hole and try to guess what should be there. This works for a watermark or a person standing still, but it falls apart the moment there is a physical interaction."

VOID approaches this differently. Instead of just filling holes, it's attempting to reimagine what researchers call a "counterfactual reality"—basically, a version of the video where that object or person never existed in the first place. Not just visually absent, but causally absent.

How VOID Actually Works

The technical architecture here is genuinely clever. VOID uses a two-pass system that separates reasoning from generation.

In the first pass, it combines a vision-language model with SAM 2 (Segment Anything Model 2) to understand the scene. While SAM 2 creates a pixel-perfect track of whatever you want to remove, the AI asks itself a deceptively simple question: if I remove this, what else changes?

Remove one domino from a falling chain, and VOID identifies which other dominoes are causally affected. It generates what the researchers call a "quad mask"—a map that tells the diffusion model not just where to erase, but where to rewrite the physics of the surrounding environment.

The second pass handles generation and refinement. A video diffusion model creates the new footage based on that quad mask. Because these models sometimes get a bit... dreamy (objects morphing, shapes losing consistency), VOID includes an optional second refinement pass using something called flow warp noise to lock shapes into place.

But here's where it gets interesting: how do you teach an AI what didn't happen?

Training on Synthetic Realities

Netflix and their collaborators at Insight couldn't exactly film a car crash and then un-crash it in real life to generate training data. Instead, they used synthetic environments like Kubri to run thousands of physics simulations.

They created paired scenarios—one version with a collision, one version where the object was never there. By showing the AI both versions of the same scene, it learned the relationship between an object's presence and its impact on the environment. It's training on counterfactuals: teaching the model to understand causality by showing it what changes when you alter the past.

That's... kind of wild when you think about it. We're training AI to imagine alternate timelines.

Testing VOID on Iconic Movie Scenes

The Better Stack team put together a custom web app to test VOID (the official repo has some significant documentation gaps, apparently) and ran it on three famous scenes: The Matrix, La La Land, and Titanic.

Results were all over the map.

The Matrix: Removing Neo from the fight scene with Morpheus left... well, Morpheus fighting a ghost. "It looks like Morpheus is fighting a ghost," the team noted. "We can see that there are some inconsistencies with the removal of the hands and other things. So, it's not perfect." Even after the second refinement pass, it still felt like Morpheus was "dancing or something."

La La Land: This one shocked them. Removing Emma Stone from the dance sequence produced what they called "almost flawless" results. "I can really believe that Ryan Gosling is just dancing by himself here," they observed. The transition where Emma Stone moved in front of Ryan Gosling was "almost seamless." Minor artifacts, but overall "a stunning result."

Titanic: Removing Leonardo DiCaprio from the iconic "I'm flying" scene gave us Kate Winslet standing alone at the ship's bow, which is... deeply sad, actually. The model did well removing Leo, but left behind a creepy disembodied hand on Kate's arm—though the team acknowledged this was a segmentation error on their end, not the model's fault. Her face also morphed slightly, creating "a bit of uncanny valley."

The wildly different results across these three tests tell us something important: this technology is highly scene-dependent. It's not that VOID sometimes works and sometimes doesn't—it's that certain types of scenes map better to its capabilities than others.

The Messy Reality of Implementation

One thing that struck me about the Better Stack walkthrough: getting VOID running is not trivial. You need cloud GPU access (they used an H100), multiple API keys (HuggingFace, Gemini), access to gated models, and you have to navigate documentation that's apparently full of holes.

Quote from their experience: "The GitHub documentation has a lot of holes and misleading information. So, to get it working correctly, there are a few things you have to watch out for."

This is the gap between research models and production-ready tools. Netflix released this as open-source, which is genuinely cool, but "open-source" doesn't mean "accessible to non-technical users" or even "accessible to developers without significant debugging."

What This Actually Means

The immediate question everyone's asking: what is Netflix planning to do with this?

The Better Stack team floated the idea of interactive narratives—imagine Netflix letting you alter video storylines based on your preferences, similar to how Black Mirror: Bandersnatch offered choose-your-own-adventure branching. Remove a character from a scene and watch the story unfold differently.

But there are... other implications here. We're looking at technology that can not only remove people from footage but rewrite the physics of scenes to make that removal plausible. In an era where we're already grappling with deepfakes and synthetic media, tools that can generate convincing counterfactual realities add another layer of complexity.

VOID is open-source, which means it's not just Netflix who gets to use it. The code is on GitHub. Anyone with the technical chops and GPU budget can run it.

The technology is impressive in its ambition—teaching AI to understand cause and effect, to reason about physical interactions, to imagine alternate realities. But like most genuinely interesting AI developments, it opens more questions than it answers. The La La Land results show what's possible when it works. The Matrix results show we're not there yet. And the implementation complexity suggests this isn't ready for mass adoption.

For now, VOID exists in that fascinating space where research meets reality—functional enough to demonstrate the concept, messy enough to reveal how much work remains.

—Yuki Okonkwo, AI & Machine Learning Correspondent