Netflix Released VOID for Free - Run It Locally and Delete Objects from Videos

Added: 2026-05-05

[00:00:01]Netflix is famous for canceling your favorite show, but this time they have released something you actually want. An open source AI model that removes objects from videos like they never existed. Let me show you a hands-on demo which is running locally on my system.

[00:00:20]So, this is the original video where we will be removing this glass with the help of this model which is called as Void.

[00:00:29]As you can see that at the moment inference is running.

[00:00:34]Let me quickly show you my VRAM consumption. So, consuming just under 13 gig of VRAM.

[00:00:40]While it removes the glass from this video, let me tell you a bit more around this model.

[00:00:46]In fact, let me play a video which is going to warm up your heart a bit. There you go.

[00:00:54]So, Void stands for video object and interaction deletion. Solves the problem that every existing video editing tool gets wrong almost. When you remove an object from a video today, you get a clean hole where it was, but everything that object touched still behaves like it's there. A ball keeps bouncing after you erase the person who kicked it. Void fixes that. Remove the person and the ball just sits on the ground because a kick never happened. It understands cause and effect and that is the selling point here for this model in my humble opinion.

[00:01:33]Not just what an object looks like, but what it does to the world around it.

[00:01:39]The way it works is surprisingly elegant and I will tell you all about it in very simple words, but let's go back here to see what is happening. There you go. So, let me play the resultant video.

[00:01:53]You see, the glass is no longer there.

[00:01:56]How interesting is that? But still that lemon is bouncing off the table. Let me play the original again.

[00:02:04]Can we play together? Yeah.

[00:02:06]No.

[00:02:09]Let me play the original again.

[00:02:12]So, you see lemon got dropped and then it is the person is using those sticks.

[00:02:18]Let's play it again.

[00:02:22]It is bouncing off the table.

[00:02:25]Now, these are the sticks which are still not perfect, but again, there is no sign of glass and the lemon drops onto the table instead of just, you know, going in the void and that is the name of the model. Let me quickly show you how you can get it installed to like I have done it and then we will check out the architecture.

[00:02:48]So, this is my Ubuntu system. I am using this GPU card Nvidia RTX A6000 with 48 GB of VRAM.

[00:02:55]And if you're looking to rent a GPU on very good price, you can also find the link to Mass Compute in video's description with a discount coupon code of 50% for range of GPUs.

[00:03:07]Okay, so let's go here. First step, let's get clone the repo and I will drop the link to it in video's description.

[00:03:15]Let's install all the prerequisites.

[00:03:18]Everything is installed. Now, what I have done, I have just put a Gradio interface on top of their demo. Just a very lightweight one. So, I'm just going to run this and this is what is going to provision this on our local system.

[00:03:34]And this is what it looks like as it runs on our local host at port 7860.

[00:03:41]Okay, in the next example, what I'm going to do, I'm just going to give it this video where a ball is rolling off the table hitting that duck and then you see that it just rolls off there in that direction.

[00:03:54]So, the mask is already there and we need to provide this a ball rolls off the table. So, the prompt which is describing the scene.

[00:04:02]And there are a few advanced parameters like inference steps, CFG, which means that how much your prompt the model follows the prompt, but I'm just going to give it the default values. Let's run the Void. It is running. While it runs, let's try to understand what exactly is happening here in this one.

[00:04:23]Now, the way Void works is surprisingly elegant. You click on the object you want gone. A vision language model then reasons about everything in the scene that was causally connected to it. What would fall, stop, or change direction without that object present. It encodes all of this into a special mask that guides the video diffusion model telling it exactly what to erase, what to rewrite, and what to leave untouched. The model then generates a physically plausible version of the scene as if the object was never there. And if the result has any shape instability in moving objects, a second refinement pass smooths that out using motion flow from the first result. The whole thing runs on a 5 billion parameter model that Netflix has released completely open source as you can see here and they also have shown some past two refinements for example, look at this. There's a basketball.

[00:05:25]There you go.

[00:05:27]How good is that? And look at this ukulele one.

[00:05:32]You see it falls off and there you go.

[00:05:36]Pretty cool.

[00:05:39]So, this is a pass two.

[00:05:41]Now, I think they have done pretty well. It's not perfect, far from perfect, but a very interesting uh I think model which they have produced. I'm not sure if they are using it any in any of their own series or not.

[00:05:58]It is still running and by the way, you can see that it takes a bit of a time.

[00:06:01]There you go.

[00:06:02]So, around 2 minutes I would say and if I show you uh the terminal, hopefully it is still there. Let me quickly just check out it.

[00:06:10]There you go.

[00:06:11]2 minutes it has taken.

[00:06:13]And I'm not sure if I mentioned it is based on that Cog video model, the base one which we already have covered on the channel, which is another amazing amazing model.

[00:06:23]Anyway, let me again first play the original.

[00:06:26]There you go.

[00:06:29]And now, let me play this resultant one. There you go. So, duck is gone. The ball just rolls off.

[00:06:36]And that is what we asked the model to do that there is no duck it removed.

[00:06:41]It has kept everything as is. Just look at um the sofa behind it, this table, sorry, [clears throat] the surface.

[00:06:51]Now, if you have to nitpick, you can say that, you know, the grain and the texture on the table's surface is not exactly the same.

[00:06:59]The model has tried its best.

[00:07:01]Um it is a bit smoother than the original one, but still I think pretty good result I would say.

[00:07:07]Let's do Let's do one final test.

[00:07:10]In the final test, I'm going to give it this video of a marshmallow desert.

[00:07:16]It is being flamed.

[00:07:20]Let's wait for this original to finish and let me play the resultant one.

[00:07:26]So, you see the flame is gone.

[00:07:28]But still, you see the marshmallows are being uh you know, they are changing the color from that side.

[00:07:37]Pretty good. Of course, you know, there is a this angle is not exactly the same and we can't see that, but still the effort is good.

[00:07:46]It's not 100% there yet, but I think already the one this first pass and this first release is quite good. Now, I have really thought hard where where Net Why Netflix built this. Maybe think about it and please share your opinion in the comments, too.

[00:08:05]I think every scene with a visible crew member, a stray prop, an unwanted logo, or a continuity mistake that would otherwise cost thousands in reshoots, maybe a model like this can just make it disappear.

[00:08:20]The most expensive part of post-production might have just become a few mouse clicks for Netflix.

[00:08:27]Uh but let me know what do you think what could be the where Netflix might be using it. That's it. Um please like the video and subscribe and consider becoming a member and please follow me on X if you're looking for AI updates.

[00:08:40]Thank you for all the support.