Netflix just dropped their first public model on Hugging Face: VOID: Video Object and Interaction Deletion

[-]

eugene20@reddit

"VOID removes objects from videos along with all interactions they induce on the scene — not just secondary effects like shadows and reflections, but physical interactions like objects falling when a person is removed. "

That is really impressive.

[-]

False-Difference4010@reddit

Pretty sure it will be used for censoring their shows in some countries

[-]

ArcadiaBunny@reddit

Second this

[-]

xienze@reddit

I bet it’s used in conjunction with a model that adds/replaces objects for the purposes of advertising (it’s always about advertising). For instance, take away the can of Pepsi sitting on the table and put a Coke in a character’s hand.

[-]

SmartCustard9944@reddit

Personalized TV show variants with personalized ads🤦‍♂️

[-]

MrAHMED42069@reddit

Uuf

[-]

TuxRuffian@reddit

Personalized TV show variants with personalized ads🤦‍♂️

This is my guess. Why does everyone on every Netflix Show have the same snack and beverage preference as me?....oh righ.

[-]

DreddKrilov@reddit

lol for sure it will be used for personalization, but no advertiser is paying to reinforce what you already prefer/do on your own. Ads are used for behavior modification, so it won't work like that.

[-]

ger868@reddit

"if people are going to fall for this" - that hesitant "if" might be the most hopeful sentiment I've seen in a while.

[-]

Vivarevo@reddit

personalized advertising too

and in some countries, removing characters, changing skin colors, items, gender etc.

[-]

Reasonable_Ad719@reddit

Plenty of cg animated films have localized elements, as it is easy to make and pleases the audience. Often it is a simple translation, like for "bakery". Now, im not sure it will serve the same purpose if a bar in NYC will have translated "bar" over it in a real movie 🤔

[-]

KadahCoba@reddit

Famous examples of element removals and swaps in media is the removal of cigarettes and smoking in many anime series that target younger audiences and that whole guns-to-radios thing from ET.

[-]

Reasonable_Ad719@reddit

It makes perfect sense - there're various legal restrictions on movies in different countries too. Although, I yet to see those removals. Prior, it was cheaper to just cut the footage out.

[-]

KadahCoba@reddit

Or chroma key recolor, like the unfortunate red (blood) to white it some shows recently. Several places have restrictions on blood and the results of workarounds are often pretty silly.

[-]

ghulamalchik@reddit

If you remove the main character what happens?

[-]

Seakawn@reddit

ever gone lucid in a dream? all the action stops, and if there're people around, they just kinda.. idle. it's creepy.

for some reason that's my first thought.

[-]

megacewl@reddit

never realized it but the few lucid dreams i’ve had, it was only me during the lucid parts

[-]

anime_forever03@reddit

Yeppp ive had a couple ones but by the time I realize its lucid i just wake up before i could do anything 🥲

[-]

megacewl@reddit

I had my first one in years several months ago, and it was the first one ever where I both caught on quick enough and managed to start trying to do things. I was even able to ‘recall’ the strategy, during the dream, that I had read about for forcing something to happen. Blew my mind that it worked.

It was SO COOL. I recommend to you to have some more of them.

[-]

pissoutmybutt@reddit

you can kinda train yourself to have lucid dreams. i read that if you can build a habit of pinching the back of one hand when you wake up, eventually it will carry into your dreams. when you pinch your hand and it isnt met with the tiny amount of pain like normal, it can trigger your brain to recognize you are not awake.

lucid dreaming is WEIRD, and not what I expected. I could never talk, or fight or do anything involving more than the most basic motor functions. usually i could “fly” in the sense i could go straight up but my only “safe” way down was to interrupt my fall with little short bursts of “flying” to slow me down. sorry if thats confusing to read, but its hard to describe since its not relatable to the anything in reality so im trying my best to describe it

[-]

senobrd@reddit

I wouldn’t recommend pinching yourself. You certainly can feel pain in a dream. Better signals are trying to turn on or off the lights (usually doesn’t work) or focusing on a clock or written text (usually looks garbled or nonsensical).

[-]

Old_Cantaloupe_6558@reddit

Just check the time twice. You know you're in a dream if the time is not consistent.

[-]

tkenben@reddit

I go lucid frequently, but if I try to stay there, I wake up, so instead I try to sit back and "watch the movie". That's more interesting anyway, because it's more dynamic and has unexpected things happen.

[-]

Frosty-Cup-8916@reddit

Lotr but only gollum

[-]

ticktockbent@reddit

Imagine the awkward silence as everyone sits around with no one to talk to

[-]

n00b001@reddit

Big bang theory without laughing track

[-]

positivitittie@reddit

I always hated seeing the laughing track.

[-]

Borkato@reddit

I’ve never seen a laughing track while watching the big bang theory. Do they visit a film studio or something?

[-]

positivitittie@reddit

Yes. That’s the one they want removed with the vision model remover.

[-]

n00b001@reddit

Urine luck:

https://youtu.be/jKS3MGriZcs

[-]

milanove@reddit

I can’t wait to see all the meme videos people will make with this technology. I wanna see Seinfeld without Jerry.

[-]

BriansRevenge@reddit

Garfield Minus Garfield: The Movie

[-]

seanthenry@reddit

That's the first place I went with it also.

[-]

isademigod@reddit

r/garfieldwithoutgarfield

[-]

LinkSea8324@reddit

Just call it Stalin

[-]

thawizard@reddit

First thing I thought as well.

[-]

hyperdynesystems@reddit

Someone needs to use this to fix Villeneuve's Dune by removing Zendaya.

[-]

EfficientWinter8592@reddit

How tf did they do that?

[-]

ghulamalchik@reddit

With examples of what would happen when something is censored vs not censored. Probably took a ton of time and effort since it's not just text, you have to show it example videos.

So they basically record the same scene thing twice.

At least this is how I would approach it.

[-]

PANIC_EXCEPTION@reddit

I just wonder how you perfectly recreate a scene with just the removal difference. I guess if you just have enough data, they don't need to be perfect? Or use photorealistic CGI instead?

[-]

Nice_Database_9684@reddit

You don't have to. You just generate the training data yourself. Film a room. Remove something. Record it again. Boom, training data.

[-]

PANIC_EXCEPTION@reddit

If it's just still images, that's easy. But you have to perfectly recreate the motion for a scene pair, and if that involves anything short of a robot, it's impossible. People can't just perfectly recreate movements. Try holding your hand under a desk lamp with no support. See how the shadow is shaky no matter how hard you try to keep it still? Now scale that up to whole body motion and irregular gait. Even facial expressions.

If the training data can handle irregularity without issue, then that's fine, but if the difference signal must be precise, then that's the question.

[-]

dvztimes@reddit

Actually I bet its better for training if it isnt a perfect recreation. That builds in flexibility.

[-]

alphaclass16@reddit

all they need is any content they own that used any removal in post. they'd have access to the plates w/and without the element in question

[-]

code-garden@reddit

Yes, they use CG, you can see the paper here https://arxiv.org/pdf/2604.02296 . They generated many physically simulated CG scenes with and without a particular object and a mask for that object in the initial scene. These are used to fine-tune a video model that already can do object removal but not the physics.

[-]

SmartCustard9944@reddit

Transformers don’t just learn sequences. Some architectures learn how to fill gaps.

[-]

Mountain-Pain1294@reddit

This is horrifying you mean with how people will misuse it

[-]

Snoo_64233@reddit

That is a whole lot of temporal understanding and cohesion the model has to deal with.

[-]

SupernovaTheGrey@reddit

IDK why this makes me think of that Stalin photo

[-]

mylAnthony@reddit

Almost sounds like april fools model 🤔

[-]

eugene20@reddit

I did have to check the date on the project page before commenting.

[-]

Mayion@reddit

"What if we remove mosaic?"

[-]

Neither-Phone-7264@reddit

why did they translate it like that

[-]

tophology@reddit

It's a meme. It's not a real fansub

[-]

tavirabon@reddit

No, this is literally a screenshot during a time where the fansub community was overly concerned with respecting the original Japanese meaning in the translations. You would have to pause the anime to read all these notes to understand what was going on because every word that didn't translate 1:1 to an English concept had its own note. They were eventually phased out because it made anime less accessible.

This became a meme because it was useless, even during such a time.

[-]

FpRhGf@reddit

This keikaku thing sounds too overly excessive and useless, but I do wish I can see more translators notes because I love reading about more context. Maybe it's because I haven't watched enough anime, but I've only across TN in English subs one time.

[-]

tavirabon@reddit

It's not as common these days (probably because official subs/dubs are more common) and ones that do will limit it to a sentence or so, but it really was a problem in the 00's. There would be paragraphs covering the entire screen multiple times per episode and most subs were hard subs back then.

They aren't too uncommon though, at least if it's pirated and you flip through the various subtitle tracks since they are rarely the default.

[-]

ArcadiaNisus@reddit

The nostalgia reading your comments hit me like a truck. I remember watching almost everything from Ani-Kraze / Shinsen-Subs.

[-]

HydraVea@reddit

I thought it was a real fan translation that became a meme.

[-]

Frosty-Cup-8916@reddit

I don't remember the group but I'm 90% sure it's a real fansub.

This one isn't a memesub though, and that does exist where someone tells either a completely different story with the subtitles or do a "abridged style" of subtitles that is still ridiculous.

[-]

Frosty-Cup-8916@reddit

Fan subs can be super weird

[-]

philmarcracken@reddit

so basically this...

[-]

Ylsid@reddit

Then you remove everything that interacts with it too

[-]

cantgetthistowork@reddit

For science?

[-]

Worried-Ad-7351@reddit

Well we now get censored content more ig. nice engine

[-]

ForestyForest@reddit

Theres this sci fi book called Star Carrier by BV Larsen where the concept of an "unperson" is introduced. It is a power that a few people have to erase a person from everyones memories and also all media.

[-]

s101c@reddit

So, censorship model to remove cigarettes from older movies?

[-]

Wiktor1975@reddit

Hopefully.

[-]

ElementNumber6@reddit

Or sponsors that don't re-pay up.

[-]

SluttyRaggedyAnn@reddit

Yup Netflix isn't generating models for the goodness of the community. They'll be using it to dynamically insert ads based on the viewer's ad interest.

[-]

kris206@reddit

that’s so dystopian! product placements that change based on advertisers and who is watching.

[-]

ElementNumber6@reddit

that's so dystopian!

Where do you think we are, exactly?

[-]

TopChard1274@reddit

in a local LLM Utopia?

[-]

fuck_cis_shit@reddit

yes. and in the long run, thermodynamics demands all utopias be local

[-]

IrisColt@reddit

You nailed it!

[-]

yaboyyoungairvent@reddit

This is already what's happening when you visit websites. If you live in USA you will get different banner ads compared to someone living in Brazil. If you visit AI subs you're going to be more likely fed ai products in the reddit ads.

[-]

s101c@reddit

This is happening with Pixar movies as well. There are multiple examples where they altered a specific scene multiple times for different markets

[-]

Poromenos@reddit

Netflix is generating models for themselves to use. They're releasing the models for the good of the community. They didn't have to release.

[-]

harpysichordist@reddit

Censorship model to remove races (light-skinned) from all movies--unless portrayed as the bad guys or stupid, of course. They've been doing it manually, so they want to automate it.

[-]

WoodCreakSeagull@reddit

Yes, of course, can't forget that white people are the real victims while Trump is sending gestapo to round up brown people.

[-]

WhateverOrElse@reddit

"If you can convince the lowest white man he's better than the best colored man, he won't notice you're picking his pocket. Hell, give him somebody to look down on, and he'll empty his pockets for you."

You have found the lowest, whiny little white man. In this thread. So far.

[-]

harpysichordist@reddit

真是可笑.

Reddit never misses opportunity to disparage white men. I am neither white nor man, yet right about little. But you are the racist.

[-]

OkDoor726@reddit

So I just came back to Reddit after 3 years being away, it's posters like you that made me leave

Yaawwwwnnn

[-]

WoodCreakSeagull@reddit

Couldn't identify anything wrong with what I said, just chiming in with "omg woke" after I replied to someone imagining anti-white racism

You were not missed

[-]

OkDoor726@reddit

Living up to the neck beard redditor I see

This place is just a sess pool of soytards

[-]

Axxhelairon@reddit

no one cares

[-]

harpysichordist@reddit

* "Gestapo" is one way to indicate how badly you're trying to distort reality. The U.S. has laws related to immigration and border crossing. Enforcement of these laws is the duty of the executive branch. If you have a problem with the laws, you should speak with the lawmakers. And your race-baiting is another indication of how badly you're trying to distort reality. Trump has offered illegal aliens thousands of dollars to leave the U.S., rather than be deported, and tell them to re-enter the country legally. He's done this multiple times. He didn't have to do this. He could have limited actions strictly to deportation. But there are people in the U.S. illegally who continue to remain in the U.S. illegally. They will have to face the consequences of their actions. But there are people, like criminals, who are upset when laws are enforced.

* Whites have been the victims of systemic racism in the U.S., yes. Ignoring it or trying to hide it doesn't make it less true. And downplaying it as only applying to the entertainment industry is another distortion you're trying to make. Is Netflix racist? Yes; explicitly so. They make race-based decisions heavily throughout their operations.

[-]

WoodCreakSeagull@reddit

"Gestapo" is one way to indicate how badly you're trying to distort reality. The U.S. has laws related to immigration and border crossing. Enforcement of these laws is the duty of the executive branch.

The same executive branch has made it repeatedly clear that the unqualified savages they employ in ICE face basically no scrutiny or accountability for their actions, as demonstrated when they lied and smeared American citizens as domestic terrorists when ICE were shown on video murdering them without cause. Not to mention all of the rapes and abuses that go on in the camps where they hold migrants. Not to mention that there is little recourse if ICE decides to just lie and grab legal citizens who just "look foreign." This is something the American government, including SCOTUS, has only decided to make much easier for them to do based on racial profiling.

Everything I just mentioned is 10000x more harmful and dangerous to non-whites than the "systemic racism" you laughably imagine.

Whites have been the victims of systemic racism in the U.S., yes.

White people enjoy by far the most systemic advantages of anyone else in the U.S, enjoying these advantages after centuries of subjugating and killing every other race of people in and out of the country, until the modern day. The country is governed right now by a white nationalist administration that is hell bent on filling concentration camps with people who speak Spanish. All of the "systemic" shit you're complaining about is literally trying to smooth over the brutality.

[-]

ZombieTesticle@reddit

Censorship model to remove races (light-skinned) from all movies

They do that with casting because the historically important characters they want to re-cast tend to be speaking parts which you couldn't handle with this.

What this is more likely for is replacing product placement on a regional basis as already mentioned, removing no longer culturally acceptable actions like smoking and probably removal of darker skinned people from the background to make shows more palatable in Asia, China especially.

The people already furiously typing would be well served to compare movie posters in the west and in China some time.

[-]

MaycombBlume@reddit

If only AI could remove the giant rock you're living under.

[-]

ticktockbent@reddit

I was thinking how amusing it would be to rewatch old movies with central plot points simply removed. Godzilla, but you remove the big lizard and everyone just stops looking panicked and goes back to their business and stuff

[-]

Long_Pomegranate2469@reddit

You can use it on that Blue girl getting railed by a thousand dudes and she'll just go and do a boring retail job.

[-]

Cultured_Alien@reddit

Except that the girl will be removed and dudes having gay party.

[-]

Borkato@reddit

Now that’s my kind of party!

[-]

MisterDalliard@reddit

Wasn't this part of an Arthur C Clarke novel?

[-]

Effective_Olive6153@reddit

People will start filming videos with generic "product" package. Once the show it published and distributed, they will be able to replace the generic "product" with targeted advertisement at point of distribution - like youtube, theater, or streaming service.

The real power is for streaming - you may have a million people watching the same show, and all of them see different targeted product placement depending on their data profile

[-]

daedalus1982@reddit

or evidence

[-]

Perfect_Twist713@reddit

Or memoryholing people and events. At least it's out in the open instead of behind closed doors.

[-]

johnfkngzoidberg@reddit

Probably great at removing ex-boyfriends from insta posts. So useless for normal people.

[-]

TechNerd10191@reddit

When it comes to LLMs, Netflix is more open-source than Anthropic.

[-]

thrownawaymane@reddit

Netflix has been posting cool open source shit for a long time. Here’s the first one I ever heard of, 10 years ago:

https://github.com/netflix/chaosmonkey

“Chaos Monkey randomly terminates virtual machine instances and containers that run inside of your production environment. Exposing engineers to failures more frequently incentivizes them to build resilient services.”

That’s my kind of party.

[-]

Zeeplankton@reddit

that's hilarious

[-]

iMakeSense@reddit

Didn't they stop using it internally at a point? I always thought it was a good idea.

[-]

thrownawaymane@reddit

That’s what I heard, no idea why. It’s not suuuuper inactive commit wise though

[-]

buttplugs4life4me@reddit

I remember when it took the company I worked for, which was a competitor to Netflix, a couple of years to actually want to do chaos engineering after massive pushback.

When we finally got a QA guy (yes, one!), his first action was to implement chaos engineering.

So his first act was to get buy-in from the higher ups for chaos engineering. There was a lot of publicity around it.

And then only his favourite time could do it while the rest of us looked on.

Shit engineering culture honestly, all the people supposed to push for that were wet noodles that bended over for the higher ups faster than a hooker.

[-]

HopePupal@reddit

remember when they invented AWS autoscaling before Amazon did? Netflix software people are not to be underestimated

[-]

TuxRuffian@reddit

Unfortunately it hasn't been updated in over 2yrs, but they also created MetaFlow (Open-Source Framework for ML, AI, & DS), although I noticed that the GH Repo says it's now maintained by Outerbounds, even though it's still under Netflix's GH Account. I wonder if the NF owns Outerbounds?🤔

[-]

Terrible-Detail-1364@reddit

yeah, theres a system I know of that still uses eureka and zuul

[-]

cantgetthistowork@reddit

First time I'm hearing about this

[-]

Seakawn@reddit

this sounds like what god does to my life in general, the difference is I don't build the resiliency.

[-]

Heavy-Focus-1964@reddit

hahahaha. what a great idea

[-]

Competitive-Ill@reddit

Fucking love chaos monkey! They introduced me to chaos engineering ❤️

[-]

pigeon57434@reddit

literally everyone is more open source than anthropic

[-]

ReachingForVega@reddit

Even OpenAI? Lol

[-]

trombolastic@reddit

well yeah, codex is open source https://github.com/openai/codex

Claude Code on the other hand just accidentally open sourced itself for a minute

[-]

ReachingForVega@reddit

The convo is about models not harnesses.

[-]

trombolastic@reddit

Anthropic has zero open models lmao

[-]

bernaferrari@reddit

OpenAI has GPT OSS, Anthropic has not

[-]

Educational_Note6910@reddit

Can't be accurate. They have opensourced Calude code twice in the last two years.

[-]

skadoodlee@reddit

Anthropic is pretty open source recently

[-]

TechNerd10191@reddit

That's why I said "AI models"

[-]

reddit-369@reddit

Some people say Anthropic doesn’t do open source.

Turns out their Claude accidentally did—

just… open-sourced itself.

[-]

Ylsid@reddit

Netflix is from the old guard of tech bros, made from people who believed that open source isn't dangerous

[-]

daniel-sousa-me@reddit

Not surprising since the whole company was built around the idea that model creators should be gatekeepers of its capabilities

[-]

Howdareme9@reddit

Not really hard tbf

[-]

HugoCortell@reddit

Too bad that it's for removing objects and keeping the background, rather than the other way around. I'd really love an AI to help with tedious greenscreen work.

I know there's a few (as in 2) models out there already, but the quality isn't great, and the set-up process is hard from user friendly.

[-]

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! Come check it out!

You've also been given a special flair for your contribution. We appreciate your post!

I am a bot and this action was performed automatically.

[-]

LengthinessHour3697@reddit

Zepp k99

[-]

PromptAfraid4598@reddit

Now we can edit video surveillance footage just like in the movies, where no one kidnapped the girl waiting for the bus.

[-]

YRUTROLLINGURSELF@reddit

What a coincidence, wow, I just dropped my new public model and it happens to be called VOID too: Viewer Offboarding via Invoice Deletion

[-]

YRUTROLLINGURSELF@reddit

wait I got a better one Value Optimization through Immediate Departure

[-]

Shockbum@reddit

Netflix/Custom-model

[-]

redditer129@reddit

Studios need to take a stand first

[-]

hugganao@reddit

it's the model made by the company that was owned by ben affleck from what I remember lol

and netflix acquired them.

[-]

MerePotato@reddit

I shiver at the thought of what this might be used for, but on the other hand, this is going to be very helpful for cleaning robotics datasets

[-]

RegisteredJustToSay@reddit

Video is a bit misleading. You have to use a 4 value mask for every frame of the video: the object, object overlap, what was affected by it, and background. Results are cool but I think they're making it sound easier and less work intensive to use than it is.

[-]

PrysmX@reddit

Someone else can take the next step and create an auto mask. Maybe they open sourced it so someone would do that for them haha.

[-]

RegisteredJustToSay@reddit

Not a bad theory. Definitely a missing part of the workflow at this moment!

[-]

NoAim_Movement@reddit

Whats the vram req

[-]

unkz0r@reddit

Neat

[-]

Grouchy-Line-4045@reddit

Wonder how long it would take to remove Jar Jar Binks from the 142 minute Attack of the Clones.

[-]

Sioluishere@reddit

:(

[-]

Lanceward@reddit

Luv me 64gb m4 max mac studio

[-]

toptier4093@reddit

Thing goes hard! Just wish I had opted for 128gb but can't complain.

[-]

Hood-Boy@reddit

AMD Strix Halo FTW. Who needs Speed anyway

[-]

EveningIncrease7579@reddit

Waiting for quantizations and kj nodes to supports it in low vram

[-]

pivotraze@reddit

I just want to Netflix or others to use some kind of AI to lip sync when changing languages. If I am watching a natively English movie in German, I want it to fix the lip sync to match. Bonus if they can make the subtitles actually match.

[-]

Sliouges@reddit

Netflix leading the way into efficient and thorough censorship. Imagine what could be done if they spent this money on ADDING objects from videos along with all interactions they induce on the scene.

[-]

Kurcide@reddit

bruh… it’s a green screen model for film making. You can’t be serious

[-]

Sliouges@reddit

It's the opposite.

[-]

Kurcide@reddit

Right, the model is meant to remove things the same way you would in film production when you have a green screen and/or actors or participants that need to be cut out

Like when someone in a green body suit is playing the role of a CGI character that hasn’t been edited in yet

[-]

Sliouges@reddit

If you remove the objects to videos along with all interactions they induce on the scene, what's the point of having a dude in a green suit at all? I have a dude in a green suit in Harry Potter, moving chairs in the pub, to simulate the chairs being moved by magical force. I use this VOID to remove the green guy AND the chairs are never moved... what's the point in that?

[-]

Kurcide@reddit

That’s exactly what you would want if someone is in a green suit or if a camera car was following the scene subject and you don’t want artifacts in the film that need to be cleaned up in post… They are only there for the actor to interact with as a representation of what will be in the final film or to get an additional camera angle.

It’s ok you don’t know anything about film making but it’s asinine to think Netflix would publicly release this with the intent of “censorship” as an open source model when it has a very clear and useful purpose in film production.

[-]

Sliouges@reddit

my phone number are code is 310...

[-]

Django_McFly@reddit

they were the person calling for a ban on all cars as soon as the first traffic accident ever took place.

[-]

mailslot@reddit

Can you imagine the ad placement opportunities? In Star Wars, every alien at the bat could be drinking Red Bulls.

[-]

Sliouges@reddit

Netflix marketing team furiously taking notes...

[-]

International-Try467@reddit

That's cool but where's Steel Ball Run Netflix?

[-]

BakaPotatoLord@reddit

What's the deal with this picture? I see it everywhere on Netflix insta comment section

[-]

International-Try467@reddit

They released Steel Ball Run, praised as one of the GOATs of manga ever written, but only one episode. With no fucking release date on the next episode or if it'll be in batches or a new episode is going to come out every single year

[-]

Frosty-Cup-8916@reddit

Not even available in Japan or a sail boat?

[-]

BakaPotatoLord@reddit

That is quite strange

[-]

International-Try467@reddit

You can even say it's... Bizarre.

[-]

Neun36@reddit

Interesting, it’s base is CogVideoX

[-]

Budget-Toe-5743@reddit

What could Netflix posibly need a model like this for? and what did they train it with? hahaha

[-]

TuxRuffian@reddit

It looks like may have used CogKit to build it on top of CogVideo. (ZhipuAI's video generation model) This is how Open-Source Software is suppose to work!

[-]

Coompa@reddit

can I use this to temove tattoos from my favorite pornstars??

[-]

Background-Ad-5398@reddit

if it gets small enough can be great for ai videos to remove the weird people that show up in otherwise good output

[-]

TurnUpThe4D3D3D3@reddit

Very cool, but V2V models are insanely computationally expensive. Maybe it’s cheaper than a VFX artist though, who knows. Very cool tech regardless.

[-]

BrianScottGregory@reddit

Those GPU Requirements. 40GB VRAM. I won't be using this any time soon with my paltry 6GB.

[-]

gurkburk76@reddit

Too bad this wasent dropped on april fools, would have been fun.

[-]

tiredgeek@reddit

As someone with kids, I could see this as a pipeline to create a "clean" version of content. Or maybe I'm the only one who has ever meticulously edited out a gratuitous scene.

[-]

ArguablyMe@reddit

You are not. We edit for ourselves too, not just for children who may be watching.

[-]

relmny@reddit

can it be run fully local? or does it require Gemini?

[-]

THEKILLFUS@reddit

Taduuumm

[-]

neuralnomad@reddit

OK PornHub, your turn…

[-]

Bolt_995@reddit

Wow, Netflix jumping into the fray.

[-]

Soft_Match5737@reddit

The interaction-aware part is what makes this actually interesting rather than just another inpainting model. Most video object removal just fills the pixels where the object was — VOID is modeling the causal chain of what that object was doing to the rest of the scene. Remove a ball bouncing off a table and the table stops vibrating. That is a fundamentally different problem than texture synthesis. It means the model has some internal representation of physical causality in the scene, not just visual appearance. Curious how it handles ambiguous cases where an object has both visible and implied interactions — like removing a person who was blocking light from reaching another surface.

[-]

the_bollo@reddit

Using this to remove pesky watermarks that jump around on videos would be interesting.

[-]

CaptainAnonymous92@reddit

So what happens if you use this to remove a main character from a live action show/movie? Do the other characters that interact with said removed character still have dialogue with them or do actions they do with them even with the character not being there anymore? Lol.

[-]

jinnyjuice@reddit

What engine (vLLM but for video) would you need to run this for Nvidia?

[-]

Live-Crab3086@reddit

just the thing for winston smiths to remove unpersons from youtube videos at the ministry of truth

[-]

seamonn@reddit

GGUF?

[-]

Nbdyhere@reddit

Holy shit 😂

This means this skit could actually come true!

https://youtu.be/68Z2ngl719Y?si=NsluLXNRZyvYiMii

[-]

quadmask_0.mp4       # 4-value mask (0=remove, 63=overlap, 127=affected, 255=keep)

[-]