What Ben Affleck's AI Company InterPositive Actually Does

0:00Ben Affleck:He's like, I don't want a text prompt model. He's like, I wanted a model that kind of attaches to a film. So it's like you have to film stuff. Then those dailies get fed into the model. And then you have a model specific for that film based on what you shot to do additional things.

0:19Ben Affleck:All right, welcome back to Da Noise. Addy, good to see you. Good to see you too. Busy week. You too. Yeah. All right. Kind of news has picked up literally this morning. with a big announcement from Netflix. Netflix has acquired Ben Affleck's AI company that I don't think anyone knew about called Interpositive. I was just telling you, I was at Netflix for lunch yesterday and I was sitting with the guys that would absolutely know about this. Didn't say a damn thing. Secret, secret. Such professionals they are. All right, so this is cool. This is interesting. So I think we've all kind of seen the clips of Ben Affleck talking about tech and AI and people were, I don't know, maybe surprised like oh wow he's really knowledgeable about this turns out more than knowledgeable he's had this company in stealth uh running i think since 2022 so i i was trying to dig into what exactly it does but the gist of it was he's building his own ai company that sort of builds models specific to each film product to create tools to help productions and filmmakers do additional things with footage that they already shot like adjust color wire Erasmus, change camera angles Stuff like that It seems like a very film language focused Model, so from what I understand from this Press statement, and just like you I had no idea this came about out of nowhere It seems like their bread and Butter is an actual model, so I'm seeing a lot of similarities With the language of the press announcements With what we covered with Brin's interview and how Asteria and Moon Valley are positioned Both as a production company and as a Technology company Yeah, reading the stuff that's really kind of reminded me a lot of stuff of Moon Valley and sort of what Bryn was talking about when we interviewed him, sort of his vision of every film would have its own custom model.

2:05Ben Affleck:And what if you could change angles? And what if you could do this? They definitely were training and building a larger model and kind of trying to make a public product. This thing is definitely bespoke individual productions. But what's interesting was reading the Netflix press release and Ben Affleck's story behind the creation of the company. He said, together with a small team of engineers, researchers and creatives, I began filming a proprietary data set on a controlled soundstage with all the familiarities of a full production. I wanted to build a workflow that captures what happens on set.

2:35Ben Affleck:Yeah. With vocabulary that matched the language, cinematographers and directors already spoke and included the kind of consistency and controls they would expect. I could be wrong here. You're pretty expert-y. but even if ben affleck or his team shoots content specifically to train a model on cinematic language and all of that stuff it would you we would need as much data as all of ben affleck's lifetime plus more like we would need millions of hours of footage in order for a video model to be robust enough to be ready for any kind of production so i'm guessing this is not that That's no. Yeah, it's not that.

3:18Ben Affleck:I don't know if you saw the video. There was also kind of an announcement video with him and Bella Bajaria. Yeah, with Elizabeth Stone, the CTO of Netflix. Yeah. And so they got into a bit more detail. But yeah, he's like, I don't want a text prompt model. He's like, I wanted a model that kind of attaches to a film. So it's like you have to film stuff. Then those dailies get fed into the model. and then you have a model specific for that film based on what you shot to do additional things like wire erasing different angles relighting people but based on the i'm assuming kind of based on the established look and baseline that you already got from actual production so it was more like so he's stuff work with real actors make stuff and then the more you shoot the more you can train a custom model for that film using their product and then you can do additional things to get more outputs specific to the film that was my impression yeah and we talked about that specific category i think that's what we you and i call vfx model or video to video model where you're not doing novel generations but you're highly modifying what is coming in as an input yeah and i mean kind of reading what it does i mean it's sort of what came to mind was like a combo of a little bit of moon valley a lot of people which is yes relighting as their bread butter and then maybe some other stuff that you could kind of do with like runway olive or cling oh three i'm gonna add one more to that i think uh the the bit that everybody in film is trying to solve is the um the gap in dynamic range quality and resolution quality so yeah a little bit of ruma luma ray 3 in there yeah i'd love to know what the outputs are these things like are they getting 4k 16-bit outputs that equate the same footage they might have shot on ari or red or whatever they shot their store stuff with.

5:09Exactly. So if Ben and the team are shooting actual footage, they're probably capturing it in raw formats. And that is really important because that's how you get the high dynamic range input into the training. And then hopefully they are making changes to the output so that it's also high dynamic range coming out.

5:28Ben Affleck:Yeah, I'm curious what the... I would assume they would have to get a high level output for this to work in a professional pipeline. Want to know something funny about Inner Positive? Yeah. So they have a LinkedIn page, and I just went in there. I was like, huh, I wonder, like, have they been posting at all? So it says it's between 11 to 50 employees, which is bigger than I thought. And they just started posting stuff two hours ago. Okay, I was going to say, is that the right one? Because I tried to find a website, and there's another company called Interpositive, but then it was like Hungry or something.

5:59Ben Affleck:I was like, oh, that's not it. And so. Yeah, yeah. Yeah. This is tagged on Netflix's press release and stuff. It's the real one. So it's as much of a mystery to the rest of the world as it is to us. Yeah, in a statement he said he had been building it in stealth. And then surprise, Netflix acquired it. Yeah, I mean, I knew about Artist Equity, which is his production company. I just had no idea he was actually building tech, which is crazy. I wonder if it was used or maybe tested on anything they did with some of the films, like the recent stuff, like The Rip. I'm going to guess maybe they tested it, not in Final Product.

6:31Just knowing how careful they are about AI usage, if there was any AI usage in the RIP, it would have had a little disclaimer or asterisk on it. So I don't think the RIP had anything to do with it. And that was also about, they started shooting it back in about a year ago. I remember some of the virtual production stuff was happening already by June or so of 2025.

6:55Ben Affleck:Yeah, the other thing I wanted to note on, because when I was mentioning what products it feels like a combo of, You know, I mentioned clinging, I mentioned runway, but I could see the appeal of this is like clinging runway. Obviously, the data sets are still questionable, depending what you and where that's going to fall legally. You know, this is clean data set trained on the stuff you're filming. I mean, I'm curious how this doesn't sound like a lot of material. So, like, what are they exactly getting? how are they using the very minimal material to like build models unless it's also using something like nano banana type ish where we had our kind of showdown and it's like oh you only need a couple reference images to get something good to make additional shots that feel like something based on a few reference images well i have a theory you want to hear it yeah let's go okay so this is my guess is what happened uh when ben affleck and matt damon were uh you know doing the artist equity deal with Netflix with the profit sharing from the revenue numbers, which is unheard of, right?

7:56I think what would happen is like, hey, we have this model. It's fundamentally different from the other models. It's commercially safe and it's cinema oriented and highly specific, but we need more data. And guess who has a lot of data, video data? Netflix. Like, well, what if you take and buy this company and then you retrain it with the Netflix data and then it's just ready to go and use for Netflix. So I think it's not yet ready, but maybe it's strategically getting there.

8:29Ben Affleck:Yeah. And also in, I don't remember if it was their statement or the Variety article, it was noted that they're not looking to turn this into a consumer-facing product or even just a pro product. It's going to be strictly for internal Netflix productions and uses inside Netflix. So to your point of training on more Netflix productions, if they have the rights to train on raw data, that makes sense. Yeah, I didn't think Netflix was in the model game, but it sounds like now they have one. And so I think some portions of the internal Netflix team as well as the interpositive team will be now dedicated to making this model usable and better over time, which is quite the effort.

9:08I mean, I'm sure the teams at Alibaba that are working on Juan or the teams at OpenAI that are working on Sora are probably 50 to 100 people, I'm guessing.

9:16Ben Affleck:Yeah, yeah, yeah. I mean, it is a small team. It also reminds me, and it seems like more of the right approach to going about the Lionsgate runway deal and the promise there when it was like the partnership. And it was like, we're going to train a Lionsgate model that's going to be like, you can, you know, kind of spit out stuff for like different Lionsgate films. And it was like, but there is no Lionsgate look. They're also different. Like, how would that even work? And then it didn't really work out. But this is like, oh, the right approach where it's like, it's not a Netflix model. It is a model for each production based on what you're doing with that production.

9:52Ben Affleck:And it just lives in that box. And just to kind of make the comparison between Lionsgate and Netflix, I think Netflix is much more tech savvy in the AI space than Lionsgate was at the time. Perhaps that's now changed. But, you know, Netflix has been in the AI ML space for over a decade, right? Like all of the algorithms and all of their search recommendations, everything is just driven by neural networks and very early forms of ML. So as a company, it's built into their DNA to be frontier adapters. And I don't really see Lionsgate like that. Lionsgate, to me, feels more like a traditional Hollywood studio, which is more about IP retention and financing and backing movies and making distribution deals and things like that.

10:33It's still really vague. Look, we have not seen a single pixel of an output from interpositive models. But the fact that Netflix now has their own capability to build and train models internally, I think this is going to open the floodgates internally for using AI-generated content to augment a lot of their productions. That's a no-brainer, right?

10:55Ben Affleck:yeah and i mean in his uh video and stuff he was talking about like you know what if we just kind of shoot more it obviously sounded like a pitch of virtual production he's like what if we just shoot with real actors and stuff in smaller spaces but and not have to deal with the logistical issues of bigger productions but still have that output that we can do yeah i mean that that that sounded like the overall kind of pitch of what he was trying to do i don't know if you saw did you see this video yeah i mean that just well this is them talking this was like some stuff that they shot before but he they had i remember they had some demo shots in here of like wire racing and generating other angles so that was sort of the initial pitch in this uh promo video but to your point i mean i think everything's just been experimental we haven't seen anything actually ship in production with that i mean shout out to brin and moon valley right like they actually have a model out that you can use today and again i don't think this would be public facing we probably wouldn't be able to use it but just to see some of the outputs on netflix shows six months from now that would be i think ideal also the netflix bar is really high for content and i would imagine that um you know if they weren't seeing initial promise of high quality model they this conversation would have not gotten very far so i'm being optimistic useful yeah if it's the instance was useful in actual productions whether it's just helping with the effects or color timing um yeah i wouldn't i wouldn't see them getting it that far so yeah I'm curious to see where this goes in the future.

12:23And we'll be covering it.

12:25Ben Affleck:You know, this could just be like another Eternauts thing where it just turns out like three months later that they make an announcement like, oh, by the way, and, you know, Stranger Things season five, we actually use this in half the shots. Right. Just tell you after the fact. That's the best way to announce AI news because you never want to do it during the week of the movie release. Then it's going to affect the box office. Because anyone's going to be like, oh, yeah, I could see what I see what the AI shot is. I see what it is. And I was like, no, you're wrong, actually. it was a different shot.

12:52Ben Affleck:Okay, next one, AI and Chroma Key. Corridor Crew built their own custom Chroma Key model. Corridor Crew, who you interviewed just, I think, last week, we all love them. I watch all their YouTube videos. They have been testing and playing around with Comfy with AI for a while now. And them being super, super highly VFX oriented, one of the biggest challenges in VFX is background replacement. And one of the challenges in background replacement is getting a really clean key. So if you have somebody with frizzy hair and they're on a green screen, how do you keep the hair particles but then get rid of the green behind them?

13:34And historically, this has been done mathematically. So you take the value of the pixel and then you sort of build edges and detect edges, feather things out, and it's very mushy. And it takes a lot of hand painting, hand work to get it right, especially when you see that he's holding this glass of water. So is the glass of water going to be a glass of green water or clear water? Because it should be clear water. So these are all the real world challenges for VFX crews today. So what they did was they have a custom model that's specifically built to tackle this problem of keying. and from the results and the videos I've seen and I've seen other people use their open source model it looks really good like it does a really good job keying where you don't need as much hand input

14:23Ben Affleck:yeah and so they trained it on like terabytes of data of keying out of green screen footage green screen stuff yeah and so you import raw green screen frame the neural network completely separates the foreground object from the green screen including highly transparent ones like motion blur, out of focus edges is the model predicts the true unmultiplied straight color of the four guard element alongside a clean linear alpha channel yeah yeah and if you go down to the tech specs it is it is very very much vfx friendly like it outputs uh exr you know high dynamic range exr i think 16 bit yeah the resolution is also vfx friendly like we're not looking at 1080p images it's going to be much higher res than that yeah i mean this is super cool they mentioned it because i asked them about there are two parts of the corridor thing if you like watch their video like a year or two ago on the sodium vapor process that's right from disney mary poppins paul de bevick yeah and that and um the process they used on mary poppins to achieve this similar sort of similar chroma keen where mary poppins was shot whatever 60 70 years ago and with the penguin sequence and she was wearing the like um she had the veil with like kind of the meshy veil and like the keen was great on that and like a lot of motion blur.

15:34Ben Affleck:And that was a different process. Great video about that to go watch. But it used a very specific prism that like Disney only had like three of them ever made. And then a few weeks ago, Disney came out with a video that was just like, oh, hey, we actually still have this prism, which the quarter crew thought it was lost to time. So I asked him about that. But then I was like, oh, hey, have you guys been doing gray screen? And then they were like, well, actually we're training this model on green screen. I guess because what this model does is it also kind of removes some of the tint of green spill.

16:04Yeah, the spill and the tinge, yeah.

16:06Ben Affleck:The subject that solves that issue. Because, yeah, we've been experimenting with gray screen because the AI roto has gotten pretty good. It could just recognize people. And then with gray screen, you don't have the spill issue as much. Right. I was surprised they went with green screen, but I guess with this model cleaning it all up, then it doesn't matter as much. Yeah, like if you... Do you see an advantage with green screen versus gray screen for... Oh, yeah, absolutely. You don't want to introduce saturation in a different tone. Avoid green screen as much as possible. Are you surprised I went with green screen then?

16:38Ben Affleck:Oh, for this? Just in general, sticking with a model for a chroma key background versus going with something that's like a gray screen. Well, yeah, I think there's a lot of green screen infrastructure in the world. We're going to continue to shoot green screen for a long time. if you go to any site, it's either painted green or blue. It costs extra to paint it gray and production is not going to pay for that. They're going to just have AI figure it out. So yeah, I think this is the transition move into AI mats and AI roto and chroma keying. So maybe the next version they release is probably going to be more suited for white, gray, and green and blue.

17:20But for now, because we're coming from the legacy world of blue and green screen, I think we're going to continue to see some of that momentum.

17:26Ben Affleck:Yeah, that's a good point. And then this model is built for that, so you don't have to shift your production around to do something gray if that's not available. Yeah, like if you go to Film Tools, one of my favorite stores, if you're in the LA area, you know what I'm talking about. It's in Burbank. I could spend hours there. And you go to the green screen section, they have the Roscoe green paint and the Roscoe blue paint, but they don't have a Roscoe gray paint yet. Like it's not standardized yet. Not yet, not yet. Joey, time to make some money off paint. Going into the millennial gray chroma key business.

17:59Dude, we should just buy paint from Home Depot, slap our denoised gray paint label on it, and make millions.

18:06Ben Affleck:It's film quality and just mark it up by four times. Yeah, exactly. Well, that's what the Roscoe paints are so freaking expensive. Is there better science behind that paint that makes it more suited for filming? I'm sure it's super diffusing. It's not just Home Depot paint. No, but I think with modern tools, if you used Home Depot paint, it could totally work just fine. I mean, yeah, most of these things would work just fine. But that's also why we're doing gray screen. If you're not doing straight up chroma key with chroma key plugins and are looking for green and need that contrast, and are relying on the AI rotoscope tools, they're all pretty good.

18:40Ben Affleck:By the way, why are we talking about green and blue screen? You and I, we come from a world of LED volume, and this was yesterday stuff. And we're like full circling back into that again. I think now it's like if you can get, you know, you look at something like Beeble, it's like if you can get a lot of the benefits of LED volumes, but without the cost, then it becomes appealing. No, you're absolutely right. To go back to some of the older technologies when they work as well as LED technologies. If you look at this, right, if you look at Corridor Key, you look at Beeble, and then you combine that with where Ben Affleck and Netflix is going.

19:22It's all heading back into a physical soundstage with minimal setup, minimal lighting. And that's going to be the future for a while until fully synthetic people and backgrounds are good enough, which may or may not happen.

19:36Ben Affleck:Yeah, we may or may not want that to an extent, especially with actors. yeah i mean i think there is still something lacking when you can't aren't able to see what you're filming like we did a shoot yesterday and a lot of it was gray screen and movie it was a weird production because it was like once it was lit then it was just literally like okay let's just move the camera here okay like action go okay we're gonna move the camera here actually go and it's just like you didn't have that like kind of beat in between setups where you can like okay let me like kind of regroup and think it's like oh we just got to keep going and keep filming but also not being able to see what we were filming like and had to be like hey everyone okay just believe that there is like a factory behind you like it'll be there don't worry it's challenging and obviously there are workflows if you level up where you can kind of build a kind of a gray box 3d world but then you just kind of get into a more complicated pipeline than we had a budget for for this production but there's ways to do it so i anyways i'm saying i think there's maybe a there's still hybrid ish workflows of like AI gen stuff but still having something so your whole production isn't shooting in a blue or gray box all day and it goes crazy not knowing what they're finding.

20:47And that's where stuff like CREA or P-Video come in where it's near real time enough where you can get enough of a sense. I don't know, maybe Beeple will have like a 10 frames per second real time version. But we definitely need stuff like that.

21:03Ben Affleck:Or something like World Labs where you can kind of spin up a 3D space pretty quickly. So at least if you're filming and moving your camera, you're moving around in a 3D space for reference. Yeah, and with that stuff I'm seeing one of the problems that is going to happen, and I saw in some of the previews of it online, is when you just arbitrarily place a 3D background, a generative background into a composite, and you look at it in real time, because the background is just completely isolated from the people, the perspective is so wrong the scale is so wrong and that's what virtual production did from the get-go is get that right because the camera is calibrated to the volume and you know the xyz coordinates and you know the units and so even if you place a shitty unreal background in there at least it scale-wise and perspective-wise looks correct versus putting any genre of ai back there you're like wait my eyes is just like it's break it's breaking like and so distracting yeah like You just don't know if your scale and everything is correct in there because it could be completely off.

22:07Ben Affleck:Right. You're not in a real 3D space where it understands everything. Exactly. All right. That was a tangent. Where were we talking about? Corridor key. So anyways, it's free and open source. Yeah. And yeah, they are going to keep improving it. And we're asking for contributions. So yeah, I'm curious to see. I'll try to try it out. But I'm curious to see what other models and stuff that leads to. Also, I feel like there's a rise in just kind of people building their own models, too. They built a model and then PewDiePie built his own LLM. Oh, I didn't know that. Yeah, that I think apparently did as well on a coding test as GPT-5 or something, but he just sort of built his own model.

22:41Ben Affleck:I'm not quite sure how. Yeah, so I don't know about on the LLM side, but certainly on the video side, what they probably did was a full-weight tune, and there is toolkits for you to do that. So one of the toolkits is AI Toolkit, a very generic name where it handles a lot of the housekeeping tasks for you so you just give it a shit ton of data and label it correctly and then you're doing a full-weight tune on a on a open source video model yeah you just you just need to have the data the training data that you want to yeah and it looks like quarter crews certainly has it and it's not just about having it in quantity but also labeling it correctly formatting it in the right resolution color space all that stuff like it's a lot of work to prep the data before it goes into training and what would it basically need for this it would be you need like hey here's stuff we shot on green screen and then here's like the the the alpha the mats we need we those kind of pairings i just need a bunch of that i do like i believe so and then yeah it broke down into like five second clips ten second clips of whatever of different green screen scenarios i think it would just more be green screen heavy than not so 90 of it could be green and blue screen footage and then 10 of it could be you know post key footage yeah okay all right last one ltx 2.3 yes sir an update from the open source model.

23:57Ben Affleck:What have you seen from this? Yeah, so LTX2 is doing really well on the open source community because you can kind of use it in comfy and you can train on it or you can fine tune it and stuff like that. So let's say we want to do a denoise podcast that's fully AI generated. You can just train it on all of our previous episodes and have a pretty good understanding of where Addy sits and where Joey sits and what they do and what they sound like and all that stuff. LTX2 was that, and I think it just came out a couple months ago. 2.3 is an update, and it's updating specifically on detail. So LTX2 was a little bit soft on certain use cases.

24:36This seems to, to me, visually, it does look a little bit sharper. And then also, I guess there is audio improvements, because it is a model, as well as vertical aspect ratio delivery.

24:47Ben Affleck:So you're thinking like TikTok delivery. So this covers that now. Yeah, and we got this demo. I mean, this cat with the fisheye lens looks pretty sharp. Like, if you look at the whiskers and stuff, that's like, as the whiskers move in and out of, yeah. See, it doesn't get lost. So like, that's a good test, actually. Cat whiskers are hard. What else we got? It's got animation. I mean, I feel like it's used a lot in animation, right? Sure. Yeah, because you can train it on a certain style. So I would imagine, yeah, animation would be up for grabs. Yeah, a lot of anime stuff. there's so much anime examples it just inundiates the other examples and they did also launch another open source tool that this one caught my attention uh ltx desktop and so this is basically an open source desktop it's to generate clips with ltx but also it's a looks like an editor yeah so it's sort of like you can edit it's like an all-in-one open source non-linear editor with the ability to instantly generate clips on your timeline.