SwitchX, Seedance 2.0, and Hollywood's IP Nightmare

0:28I did a test with us. a few days ago and got a sneak peek at it, but it was officially released today. This thing feels way more practical in the AI world for actual filmmaking than most of the other stuff that we've seen. What's it called? Let me break it down. It is called Switch X from Beeble. I also want to clarify, because I've had this confusion as I've talked about Beeble to other people. This is Beeble, the AI VFX tool set. This is not Beeble the artist. I've been like, oh, you know Beeble? And they're like, yeah, I know Beeble. He's Art Basel. He does this stuff. Like, different people.

1:01So Beeple's main thing was this really cool kind of, like, you give it a video, it breaks it down, and it does this AI relighting, and it gives you all the elements that you could then bring into Blender or Nuke or whatever your compositing tool is. And it was kind of this good middle layer. But the problem was you still – they had an online tool, and it was okay, but you still had to know compositing and, like, be good at VFX. It wasn't sort of a one-stop shop. by this new tool, SwitchX, is basically video-to-video transformation, but it does not mess up your source video, and it does not mess up your people.

1:39It is the best preservation that I've seen in their demos and a couple of tests I ran that keeps what you want and changes everything else around you. Kind of like that McMuppets workflow that I always keep talking about. I'm like, that thing was cool, but you got to run comfy, and it's a whole crazy thing. this is like and you have to mask manually and all yeah this one a mask but let me i can break it down because i have a i have a demo up but uh and some clips nice okay but basically it's got this uh you know you got your source footage woman walking in the woods it'll also it automatically do the alpha and it runs a lot of the kind of biebel secret sauce stuff under the hood to like mask and relight your subject lighting engine you give it a reference image so like your first frame you style you use not a banana whatever you want to style what you want the scene to look like and then the video generation uses that first frame reference and the masking to generate everything else but preserves it will relight your subject so they blend in but it'll preserve everything it'll preserve them so let me go to well it's doing much more than relighting Joey it's changing costume it's adding footsteps in the snow it's putting you into that background so like it's doing a lot of heavy lifting yeah everything else around the person but even the subject itself if you only go to the Hoon's video.

2:54So this was their launch kind of demo video. And so what we're seeing in the big screen is the final output. So Hoon's just sitting in his office. Yeah, so that lighting feels really accurate to me. Like if you go back and just, yeah, I mean, we could just keep playing. Yeah, he's getting relit based on the environment. I mean, the relighting you're doing is based on your reference image. So you give it first frame, have it modify your reference image with the new lighting that you want. and then it honors that. Yeah, so much of it is white balance. Like they're getting the white balance right.

3:28And also it is highly aware of where the strongest light in the scene is. And he is sometimes underlit because there is no light in front of him, which is accurate. And it could just be a matter of getting your source image better or having that correct. Yeah, some of them are overly contrasty. Like the last example in the NBA gym, I want to see the relighting here. this is the relight game of thrones throne yeah yeah this is see that top down light yeah all that shadow under his eyes and nose that's nice yeah i mean the main thing for me is if you it's hard to see on the smaller screen but the source video and like what they look like doesn't get messed up because like we've had video to video with runway uh olive or uh even cling and cling is like cling 01 has been sort of the best i've seen but it still messes with the face a little bit you get a little bit uncanny valley this does not mess with the face we're so like human beings are so sensitive to facial distortion like you stop looking like a person the minute it's like 99 yeah just one little like that one percent makes all the difference yeah or even that point one percent where something's like a little bit weirder and you're just like feels something something feels off so uh i did a test with us and i could run another test and so through the Oh, the Cyberpunk alley?

4:46Dude, this is like the most cliche example. I love it. So basically the process, you give it a input clip, and the limitations right now are... Oh, so now we're looking at everything web-based, too. Like, no nuke needed. Exactly. And that's like the video, the ongoing saga video that started with Beeple that we were doing for the channel here that'll be out hopefully in a few weeks, now that we've hopefully finished filming. But it was supposed to be, it started off as like recreating the one battle after another chasing uh with an iphone and we shot some stuff in biebel at their uh studio a biebel slash partnership studio in like november and part of the delay besides the holidays and stuff was needing like a new compositor to put it together for like that last mile to finish it and then in between then conrad is like that test clip has like unearthed a lot of issues and also helped develop like four different new products like an unreal engine plugin from people and was also like what are the test cases to develop this switch x thing where it's just like oh what if you don't need to have nuke and all that stuff and you could just kind of do this here so now that we have this new workflow the the you know the video get done faster i i mean look i'm i'm wearing a virtual production hoodie right now biebel has start at the beginning of the end for LED volume.

6:15You still think it's going to be completely replaced? I still feel like, if anything, AI can be another boon to virtual production workflows at a higher level if it just drastically cuts down on VAD and stuff. You're talking about two different things. Virtual production, the big umbrella, I think there is camera scouting, location scouting, you know, all that, like the, what's that company we talk about? You mean just like, like just building it? Like VCAM stuff. Yeah, just pre-visiting. Like a World Labs kind of thing? Yeah, yeah. So I think that stuff will definitely get supercharged by AI and get better over time.

6:55What I think is the end of the road is specifically a very niche tunnel of virtual production, which is LED volumes, because it's so expensive. It's so hardware intensive. It requires so many people at the same time to run a stage. It's physical production, which by nature is so expensive. So the two big promises of LED volumes was accurate lighting and background replacement. And Beeple's doing both right now. Okay, so maybe not to the level of quality that we need for a 6K, 8K sensor. It's not there yet. The stuff does look soft. I'll give them that. But this is the beginning of the end for that specific avenue of virtual production.

7:38In my humble opinion, I helped build the industry. I sold many LED volumes. I built them by hand. And every time I was doing that, what I was wondering was, what is the ROI on this? How long does it take to recuperate this? And we've done projections. We've done forecasting. Within 10 shoots, the studio pays for itself. but now if this is running on the web and it's like a$1 API call you can run this a thousand times to get the result that you need. When it comes to a virtual production cost how much of the breakdown is just the actual day on the stage versus how much is building out all of your 3D assets and worlds and VAT and all that stuff?

8:22Great question. So my numbers are maybe just a year or two outdated But on average, if you're going to commit to an Unreal Engine scene, you're looking at anywhere from 50 ,000 to 100 ,000. That's two to three artists over a few weeks. That's for tier one, maybe even high-end tier two film, right? If you're doing plates, it's obviously going to be a lot cheaper. Studio rental rates are, so for the big badass ones like, you know, Nant, Australia, those places like that, you're looking at high five figures to low six figures a day and then for the mid-size stage you're looking at you know anywhere from 10 000 to 50 000 a day and that's without you hiring the crew right then a cinematographer has to come in the grip has to come in the acs yeah usual producer yeah it's physical production right like you're putting people in a stage and their safety requirements and all that stuff.

9:19I know for a fact that was even cheaper than taking people on location. A perfect example is Society of the Snow, that Netflix movie. It took place in the top of the Andes Mountains. How expensive is it to take a crew to the top of the Andes Mountains or even any mountain with snow for that matter? Incredibly expensive. but then you bring the snow into an LED volume. Yes, that makes sense. It's way cheaper, way more controllable. The weather's not an issue. I get it. But now if you shoot that crew on a green screen and then you can replace that background with Bebel with the snow background and then relight the actors to like a diffuse cloud snow environment, you have just eliminated the need for LED volume work in my opinion.

10:10Yeah, I get what you're saying. I think it'll probably still be two branches of like, yeah maybe tier one top tier stuff some led volume used they want to have that in the filming sense like have that feedback and have that environment while they're filming but it is like a premium experience versus you could film in a green or blue stage uh which is sort of what the b-bull vpx stage over the valley is kind of this big blue box or i mean look also to like throw stuff out there even more. We're going to do a shoot and we're going to do gray screen. And I've been hearing a lot more gray screen uses.

10:47Gray screen is really good for AI. Yeah, just because if you're relying on the AI rotoscoping tools have gotten so good that gray is good and then gray, you don't have the spill issues that you get with green screen. It's a neutral spill. Saves a lot of the headaches. I'm going to test it out first and then we're going to shoot it before I commit to that. But yeah, it could just be a soft gray box and this workflow. Yeah, I'm not saying the LED volumes will go away overnight. I think the driving plate stuff absolutely makes sense. That is the number one use case. You put a car in a volume, all the reflections are free, and there's a whole plate workflow behind it.

11:24It's come along so far. I'm saying historically, we went from shooting on location to green-blue screen to LED volume, and now we're going to come back down to gray screen and then sooner or later we're going to go back down to shooting anywhere. You won't even need a gray screen with the depth map innovation that's happening now. Yeah, that's even some of their showcases. None of these are all shot in some environment. This dude's got a door behind him and it looks like he's in a street. Podcasters, this girl is always on their promo videos. Stock, yeah, it looks like stock photos. All right, back to us.

12:03Sorry for the segue, folks. the workflow of how what you could do with this anyways so what are the limitations well speaking of how this is not going to completely change everything just yet i'll upload a new clip so you upload a video clip so current limitation 240 frames so about 10 seconds if you're at 24 frames a second one nice thing and this not a lot of editors have this is uh it gives you the option to trim the clip which i really like that because i've been on other apps where you upload the clip and then just gives you an error like your clip's too long that saved them so much gpu i bet yeah So then you upload your video and then it runs this detecting subject feature.

12:40And it automatically does a pretty good job masking me. But you could come in here and add additional masking. That's good masking. Yeah. Yeah. And it shows you the preview of the mask. Or you could also just use the entire image if you want to do something like keep everything the same. but like modify one specific thing like uh change the clothing on your person you know so you don't want to mess around with the entire scene you just want to keep them as is but add something to them or modify something with that and then they have nano banana built in but you could just take your first frame and you know photoshop or whatever modify what you want the first frame to look like you got any got any ideas of oh uh background ideas yeah background ideas let's go roman Colosseum.

13:26Like an radiator. With a top... I like the top-down sunlight thing. I think it'll play well. Noon lighting, top-down sunlight. Top-down sunlight. That'll take a little bit. The limitations... We're still in the AI world, so it's roughly 10-second clip cap plus 1080 max output at 8-bit RGB. So you're not going to... If you were shooting on... something you know a red black magic black magic raw you're gonna that's all gonna be super compressed and they're they've said like the ideal workflow oh this chain you know me uh you gotta keep the framing yeah yeah i gotta these guys should talk to topaz because if if you pair this with really good upscaling like that's a lot of the use cases for vfx that covers a big portion of it yeah i mean they would still say you know they're in talking to kairo and stuff like this workflow, you know, they're not saying this is a VFX pipeline replacement right now.

14:27This is like for web fun stuff because their other pipeline that they started before is specifically designed for high quality footage. We'll create the lighting maps, the PBRs and all that stuff. And then you composite that with your original footage and Duke or wherever with their variety of plugins. So like they kind of have that pipeline already, but, um, okay. Yeah. You got a hard shadow on you. Let's go with that. See how it behaves. And then you got your prompt and it kind of just auto-generates where you can modify if you really want to. But I'm going to leave it. And then, yeah, you got your resolution option, so I'll just leave it at 1440.

15:00I wonder if Beeble is doing any work for any of the big studios. I mean, I'm sure every studio or VFX house is using Beeble Studio in their pipeline. Right, whoever's using it. I mean, I've also heard people just say they like it, not even for the relighting stuff, but just that Beeble's keyer, like rotoscoping, is just really good as well. It's just good. Wow. Yeah. because I've heard some VFX artists say they don't even need the relighting they just like it to make keys this was a clip I did before and so you could see yeah I mean just knowing how we light our studio this hasn't really impacted our lighting at all I mean I still have the highlight on the top of my head from the light that I put here see this thing the only thing it did was really add that purple tinge on the table from that neon light and that could have also just been my like this was the reference image I gave it So like I could have also just not have done as great a job at the reference image to really.

15:53Oh, yeah. Yeah. I think that's it. You know, so it really comes down to like how good that source images that you give it of like what you want it to make. And it honors that. Oh, there's a compare button. Oh, that's cool. Oh, wow. That's even that really helps out the effects. Yeah, there's a huge white balance. Yeah. Yeah. I mean, I'm just looking at our likeness and like you like I mean, you're you. Yeah. Look at the hand. Look at your hand. and the movement. And your hand wasn't even in the shot when the first frame started. That's true. Feels like a little gap sometimes, but I mean, that's kind of crazy.

16:27That does help sell the effect. That's a good marketing tool. I was like, it's not doing much. Oh, it's doing a lot. Good job, Hoon. Hoon and the team, you really put this great marketing effect just for us to demo before and after. The wipes, man. The wipes will get you. Oh, cool. This was pretty fast. Oh, it had a rogue frame of you. I should have trimmed that. That was quick time being annoying at trimming clips. This is a shorter clip. I mean, like, as you're moving your face, that shadow is pretty static. Like, it's behaving accurately. Yeah. Like, half of your face in that shadow. Even, like, the shadow of the wire on your left shoulder is top down.

17:08Yeah. Dang, dude. Well, that shadow is in the source. Oh, it is? Okay. Oh, yeah. But it did amplify. It did make it more contrast. The entire right half of your body is darker from the actual shadow. Yeah. And I mean, look, my face. It didn't mess up my face, but my face is in shadow. Even darkened the laptop down, like the laptop's crushing. Yeah. Yeah. It's annoying with this trimmed clip. Yeah. So I'm excited to mess around with this more. Me too. Yeah. How do I get access to it? It is publicly rolled out into beta. All you need is any paid Beeble plan. I think it starts at like 20 a month.

17:45Okay. And usually Bebel runs on credits, but I think right now that showing that this is free during the beta. So as long as you sign up, you can start generating stuff. This also might be a secret way to use Nano Banana and not use credits right now. Sorry guys. Oh, for Nano Banana, I just do the free pick infinite thing. Yeah, I know. We're on the free pick infinite chain. All right, cool. Yeah, I'm excited about this one. Very, very cool. Hey, hats off, people. I know this is There's a lot of demanding people that are looking at this nitpicking, but you guys are really paving the way for AI VFX.

18:22The coolest video-to-video without messing up your source video that I've ever seen. All right, next one. C-Dance 2.0. C-Dance 2, yes. How was your feed this week? You know, I'm starting to get a little inundiated by best video model ever. Every week. Hollywood's cooked. Yeah, because when Kling 3 came out two weeks ago, that was the best ever. And before that, we were raving about Kling 01, and before that, 126. It's hard to tell. I think somebody in our comments on the last video was like, you need a metric. I forget what kind of metric, but we need to have a more consistent ladder where we rank these models.

19:06And for that, we need to test them. Which is hard when you don't have access to the models. This one is like the weirdest rollout. This felt reminded me of Nana Banana Pro again, or Nana Banana again, because like I kept seeing clips online since like, I don't know, last weekend and now it's Friday. But it was clips from the most random like accounts or products that I'd never heard of. And I was going to ByteDance's website. I'm like, did they officially release this? And they hadn't. And it was like days where I kept seeing C Dance 2 chatter. if you there was another there was like a cdance.ai and it was you know i remember the and yeah so it was weird that i think some people had some companies maybe had early access and leaked it or gave it to some influencers and stuff regardless now it is officially out at least announced by by chance i still haven't seen it on foul or uh free pick yet but i'm sure that'll come That aside, it's more like been chatter, but I have not fully believed anything I've seen because there was not an official release.

20:11Now there is. So now we can break down what is actually officially in this new model. Yeah. Honestly, we've seen a lot of the demos, but it feels similar to Cling v3. Well, it's catching up, right? These guys have to come up to a certain level, and then one will up the other. So I think the industry as a whole is catching up to VO. And Kling has caught up to VO. Now it's by dance, see dance's turn to catch up. And then when everybody's caught up, then somebody will surpass them. So there's like a rung in the ladder. I mean, this feel, let's see. I mean, let's break it down. Maybe there's something that is, you know, level up from Kling.

20:52But, you know, it's got audio and video generation in the clips. It has, you could give it a video input now. supports images, audios, and videos as references. So we see this kind of reference video of two guys fighting and reference images of... Oh my god, people are going to kill me if I don't... Whatever the fighter guys are. And then it turns into... Is it Street Fighter? Yeah, it's not actually... Street Fighter-esque. Yeah. Why is it that the anime examples are always spot on? Like, they nail those. It just sucks for the anime industry that, like, but demos are always so convincing give me some stats man we need numbers what do you think is it still a 1080p model or are we going a little bit higher here I would assume 4k but let's see I don't know if it says 1080 but let me see if that could be a total BS I need that paper am I missing something yeah I mean what you're going through is I think what everybody's going through with C-Dense 2 is like it's very mysterious and vague at the moment but there is a illusion of high quality that exists in the internet.

21:59I don't know, we'll go off whatever Higgs field's advertising, which, take that with a grain of salt. One-click video recreation. Yeah, I mean, we'll talk about this in a second, because most of the clips have also been... Oh, look at that main change rip-off, man. Yeah, and at least this doesn't have the characters. Somebody just got shot in the head. Feed up to 12 assets at once. Okay, so we know images, video, audio, clips, and text. Frame level precision. What does that mean? I think that's the same thing that Kling has in the other models of, like, multi-shot generation of the same video so you know the same idea we're making cdance 1.5 or even one was the one where they that was an accidental figure out that if you prompt multiple shots in your prompt you can generate like your 10 second video but have multiple shots yes and it would be in the same generation so it's consistent cling three does that yeah cling three does that and then obviously because cdance was the first to do it they do it but I think even more consistent.

22:54Multi-camera storytelling, generate new storylines, or extend existing videos with natural shot connections. I don't know what's the max generation, max length. I'm just looking at temporal consistency as cameras move around, pan around. That seems pretty solid on those demos. Yeah, I mean, the stuff is the quality is there. I feel like a lot of the models now it's like that is the bar between cling and via like everything's looking more realistic and sharper and less like weird stuff happening we're so close we're so close to a model that i think will come out this year and i predicted it that will be our quote-unquote vfx model like we're actually going to use it for film yeah all right so uh this is from google's summary so also take this with a grain of salt but it's saying 4k resolution 15 to 20 second duration that could be total BS because this is an AI summary.

23:52But let me see if it's on follow yet. I don't think so. I checked. All right. Sorry. I wish we usually I like to have more details in this and I feel like we're scrounging. But it's been such a weird this is a sort of summarize. Let's just cut to the Tom Cruise thing. Come on. Clicks popping up and then trying to figure out what does this model actually do and how are people using this model. Okay. Tom Cruise. Tom Cruise. This clip has sort of sent everyone into a tizzy. So yeah. I saw this on Variety. Variety covers this. You know as like IP infringement thing obviously what what ip here the people the now i'm joking the dudes they get an action film with extremely detailed look-alikes of i gotta say that's really good like uh the likeness of both of those guys at their current age that's spot on even um if you look at like the hand contact where like they're blocking each other and like the arms are making contact that stuff is accurate yeah i mean the clip look good i've seen a lot of lord of the ring stuff this one this one actually was kind of funny where they they just fly to mordor and drop the ring in the bottom you know we joked about that like 20 years ago that is yeah that has always been that was always that was always like wait if the eagles could rescue them why could the eagles just like you could have saved nine hours of film yeah a live action adaptation on attack okay so the bottom is generated live action from the top which is anime we're looking at a spaceship over some sort of land.

25:16A glimps. A zeppelin. Basically turning an anime film into a realistic looking live action. It doesn't work. It doesn't work at all. Yeah. Because anime is so limited in animation. Let me tell you what Twitter loves. Twitter loves when people take a beloved anime film and try to make it live action. The gist is the bulk of the stuff that has been popping up from Sea Dance Online has been hugely infringing on using existing IP and real people. And that's led to the deadline. Sorry, not variety, deadline. Yeah, that's it. Yeah, this is what I was looking at. So yeah, MPA calls on TikTok owner ByteDance to curb new AI model that created Tom Cruise versus Brad Pitt deepfake.

25:57So yeah, this has been the recurring theme of the Chinese models. Different view on copyright. Takes a lot of stuff that is copyright protected or likenesses and easy to generate. I mean, I feel like this clip got a lot of attention, but I feel like most of these models can do this anyways. Maybe this clip just kind of broke out of the AI bubble, but this is not new of what you're able to do. If a model is training on publicly available data, it has the awareness of Tom Cruise and Brad Pitt already. They're all over the internet. So in order for them to not generate it, they actually have to put filters or prevention mechanisms after inference.

26:39Yeah, it knows it. It just has to block people from doing it. Exactly. It's built into the model already, so you can't fight them on, well, don't use the training data. Well, there is 10 million images of Tom Cruise on the internet, and it's really hard not to. I think the studio should go after the post-inference controls. Finding a way to standardize that and make that very transparent rather than the training data. Yeah, and I still think that's where things will go. And it's like, yeah, first off, the stuff's already there. It's trained in the model. It's there. But yeah, you could add blockers and restrictions on what people can generate and do, which is, I think, what they're asking for.

27:19At a prompt level, if you detect the phrase Tom Cruise, you can say, sorry, that's not allowed or something like that. Yeah. I mean, we had this issue last week when we were doing the Genie, and we were like, Dodge Charger car. And it was giving us an error, but it wasn't saying why. and then when we stopped using dodge charger and then it generated and it was like i was probably just blocking like the specific copyright word or trademark right but yeah these models know stuff i mean the one of the projects we're working on there's a character named mickey and sometimes the name ends up in the prompts and then the outputs we get is mickey mouse like holding holding shotguns and like good so i want to see the bloopers reel on that yeah so yeah the bottles the bottles know this stuff and you know you could well that was one of our predictions too that this year that would get sorted as fair use or not or more clarity on that but yeah as we've said before the focus should be on the outputs and yeah i mean like the bigger picture i wonder is like three to five years from now when these models are really good i'm talking like discern indiscernible from reality good and anybody can generate tom cruise brad pitt whatever is tom cruise and brad pitt actually going to care or they're going to embrace the fact that people are bringing them back into the um cultural sphere with like a cool meme or something viral i think the issue is when you start having videos of them doing stuff that they don't approve of or like or you know is disparaging and and then it's just so you can't tell that it's fake harmful like this this you know fighting on the roof thing like you know it's all like haha what's the big deal like whatever but when it's like you know whatever i don't want to say like just bad examples like like stuff tom cruise visiting the epstein island or something right or yeah or yeah like stuff like that that wasn't me or or even uh i mean remember that crazy brad pitt scam where that woman was thought she was having like a dm relationship with brad pitt and then he was in the hospital and he like needed money and basically like wiped out her whole account but the images from that were like did you ever see the i think we don't i don't this is this is weird okay let's let's take a look brad pitt hospital okay this post article says ai generated brad pitt in hospital i ai generated is very generous this looked like really crappy photoshop uh french woman duped out of 850 000 by scammer posing as ai generated brad pitt and this is what her name was Ant.

29:54Oh, that is bad Photoshop. It's not even AI. Yeah. She was getting sent pictures of these and believed that Brad Pitt was in the hospital and needed her help and that she was in a relationship with Brad Pitt. This is a bad example. This is a bad quality example. Just imagine when it becomes so good and then you just have... Well, I was just thinking, just put this through Nana Banana and have it upgraded. It'll be so much better today. That's the risk. You could just make these photorealistic scams and scam people out of money and they think they're talking to the real person. Just be suspicious of everything online if you can, actually.

30:31I don't even know. If you can't see them in real person? I was going to say, if you can't, FaceTime with them. But soon, CREA real-time AI is going to be so good, you could probably have some person transform in real-time into Brad Pitt or whatever. All right. Darren Aronofsky. This one, you know, happened a few weeks ago. We never really talked about it. So Aronofsky's Primordial Soup company, which we talked about last year when they sort of announced a partnership with Google DeepMind and Veo. One of their first projects out is this weekly-ish series with Time Magazine where they're kind of dropping short episodes about the Revolutionary War, the American Revolutionary War, leading up to Independence Day and our 250th anniversary.

31:15the first few episodes are out reactions to the look of them has been it sucked it sucked we're gonna we're gonna watch it together again we can play some clips or something i don't know okay okay all right as you're pulling it up did darinovsky actually review this thing and sign off on it or it was just one of their people that just or something yeah i do think that from what they've said about the process you know they're like it was sag voice actors with the voice performance, but then everything else was AI generated. That's so irrelevant if you look at the actual visual choices and directorial choices.

31:56It just has this synthetic sheen all over it, which is so distracting from what... You're talking about the skin tones and everything right here. Everything is uniform and cut and paste and very textbook. High school textbook. To me, it feels more like, well, like for one, because AI models keep changing so fast, like what time frame was this made? Just with the model availability. And then the bigger thing is like this is part of the challenge. Like Primordial Soup has this partnership with Google Demind. They're locked, I'm assuming, locked into only using Google products, which like Vio is great.

32:36but in the landscape of everything especially something like this there are no motion transfer human performances video to video anything in vo the other big limitation of vo is the inability to customize uh look and feel right i mean you could throw in a month we tried this but like if you want consistent look and feel across the entire film you probably better off building a laura or yeah if you wanted to i mean or just get good at prompting i mean i'm just thinking more of for something that's so heavy on character speaking and then it's like, cool, yeah, you're using the voice actors, but what would probably be even better is if you had their actual performance and then you could motion transfer that.

33:17There is no model from Google that does that right now. Yeah, and also there's so many close-up shots and the whole thing just feels like it was just kind of cobbled together with what you can do with AI today. I'm guessing the workflow was they recorded the performances and then prompting, text prompt, text and image prompt with what you need them to say and then just keep regenerating until you can get an output that has the AI mouth moving that syncs up convincingly enough with the audio you recorded and then some creative editing to cut around when it falls apart. Yeah, it doesn't look like they even did color grading for that matter.