Genie 3, ElevenLabs Music, Lightcraft Spark & More! | This Week in AI for Filmmakers

Google DeepMind's Genie 3 stole the show this week with its persistent world generation, outshining OpenAI's underwhelming GPT-5 release. Join Joey and Addy as they break down how Genie 3 creates navigable AI worlds with accurate physics and memory—allowing users to paint a wall, walk away, and return to find their changes intact. Plus: ElevenLabs' music generator, Houdini 21's VFX updates, Lightcraft's new tools for indie filmmakers, and the ongoing ethics debate around Grok Imagine's controversial NSFW capabilities.

---

The views and opinions expressed in this podcast are the personal views of the hosts and do not necessarily reflect the views or positions of their respective employers or organizations. This show is independently produced by VP Land without the use of any outside company resources, confidential information, or affiliations.

Transcript

: 0:04

Welcome back to Denoised. It is going to be our Friday, maybe Saturday, weekly roundup of AI roundup. All the new AI tools and updates in the world of filmmaking. It was a busy week by Tuesday. There a bunch of updates. And then we had the GPT-5 release yesterday. Yeah, we were stoked about that. I was stoked about the, well, everyone was kind of stoked about the announcement and then, you know, the air left the excitement. So we'll get to the actually most exciting update, I think, this week, which is not GPT-5, which was Genie 3 from Google DeepMind. Yeah. So this is... an update to their world model, but the demos from this are absolutely wild. Google dropped another banger. This one. Yeah, this one is great. This one is a either from a text prompt or I've seen demos where people can give it an input video or image and it creates a complete AI world that you can move around in. And the physics, the reflection, the behavior feels like you're in a real world. I mean, like, this is like probably the biggest world model. That's crazy release that we've seen so far. Are you able to then take that world into other things? I don't think so yet. And also this is still like in a research preview. So we've just seen clips from it and they've, or releases from people that have access, either researchers or they said some creators and you could, you could run it for, it runs at 720p at 24 frames per second. It will run for a few minutes because of the memory issue of how much the world could generate. like, as it's generally in the world, the world is persistent. So there's a wild demo of this. Paint roller. This is all in Genie 3. So painting the wall. But that wall doesn't move. No, exists. It's all fake. The entire thing is generated. Wow. So then you look at it. with the keyboard and moving. And then it turned back. the paint is still there. It's still there. Oh my god. Oh my god. That's freaky. That's why there's a couple minute limit. this is, I mean, this is a bit, in Genie 2, I think the runtime was like 20 seconds. now we're two, three minutes. uh, uh. Give me some VR goggles. Put me in there. So mean, yeah, models are a thing that everyone, all the AI companies are trying to generate and have talked about. Runway's talked about. Luma's talked about. Persistent world models. Persistent world models. I mean, there's obviously the M &E implications, gaming implications, generating a game on demand. But the real reason they want this is this is a big tie-in to AGI, because it's a way for the AI to just understand how the physical world. actually exists and behaves physics and everything. Yeah. And then also autonomous vehicles, robots, robots, the robots need to train themselves on a world. Yeah. They need, if they want to operate in this world, they need to have their own understanding of how this world exists and functions. Right. And that's what the world models do. And it's, it feels like Westworld because it does feel like West because it's like, Hey, it's all fun and games. We're doing this so you can make video games or movies. But like the companies are really doing this. Like videos are the way to train it, but the money and where they really want to go is Robotics the money is absolutely bottomless digital twins You're also talking about military simulations and battlefields and that is billions of billions of dollars in contract I'm sure Google will not hesitate to take that paycheck and then yeah the robotics training we talked about and then on the industrial training side, so Let's say you're building Airplane or something if you're Boeing or your Airbus or somebody like that the ability for you to generate a warehouse with all the parts, all the gizmos, all the assembly line stuff for that airplane and then train a robot into it or train human into that environment. So then quickly ramp them up to then produce. The digital twin thing. And Nvidia always keeps talking about this in their keynotes and stuff of like, before you build the $100 million factory, build it digitally and you can make sure it all works and functions properly. Absolutely. Yeah, huge implications. Implications are huge. And then we just talked about Hunyuan's world model last time. Yeah, mean, I think that's the one. Yeah, the onion roll model I read was you can make me take a take a couple steps. Yeah, it's like not a wide lens, just like this one. This one you can move through this world. I'm not even walking. I've seen drone versions where like people are doing aerial moves through. Yeah, that the paint roller thing is mind boggling because it not only has like a flat surface, a 3D surface, if you will. but then it memorizes what you did to it. And then you can come back to it in the exact same place. Here, this one, they took a Veo 3 output of a drone shot and used that as the input. So the beginning part is Veo 3, and that was the input into Genie 3. And then now this is Genie 3, and they're like flying and navigating through this world. It kind of reminds me of the early days of Gaussian splats, where we were doing this with like a very crude low resolution version of this. we were blowing minds with that, right? You remember? when somebody had taken a drone footage of the Grand Canyon and then in Gaussian Splat recreated it and then now you can walk through the Grand Canyon just based off of one minute video. They did one of the entire Grand Canyon or just like a section of it? Just a section of it. But it was convincing enough and I think to a lot of M&E folks, especially the virtual production folks, it was like, oh, well now I can get parallax on the Grand Canyon if I put that Gaussian Splat up there. Right, when they realized, oh, I can take this, this in that application. Yeah, I mean, I think the first time I was exposed was like even just a simple scan of like from the the bad decision studio guys. They did a scan of like a building and they did a scan of like a totem pole. Yes. And brought it with a drone and brought that in. I remember that. That was first time and I was like, oh, that's pretty sweet. Yeah. And you were able to then bring it into Unreal Engine and then compound to it other things in engine like actual 3D assets on top of that, which is really cool. Yeah. So that's why I originally asked, can this be exported? Because like How many people want this in their pack? Right. course. this talks about what we've talked about with like having a, you know, building out a virtual space or virtual production, but like an environment, a virtual environment that's consistent and you can have different angles. would say at the very least a 3d file that is not 4d. So just like a static photogrammetry of that environment with a basic texture map for starters. I mean, it's like, what's the say you like, you can't just walk around the space and you know, it's a base of a video of a 3D space, they could just do the same training of a mesh or model as if you were walking around with the camera or doing like a, the other big use case is autonomous vehicles. They need a world to train in. exactly. Yeah. Holy crap. So yeah, I think this was probably the biggest, most exciting update this week. Very cool. Pretty wild. Okay. But then we had GPT-5 update, which the keynote was yesterday. Yeah. And Sam Altman was on Cleo Abrams like right after, I mean, they prerecorded it. was it? Yeah. She dropped that episode. I didn't see that interview. And Cleo was like, why don't we talk about the near future and long-term future of where you think AI will head, and obviously Sam Altman's view of the world is way the heck out there. Yeah, I bet. So look, I mean, the big updates, I think, probably the biggest, most user experience update is Previously, and we've made fun of, and everyone's made fun of the like 20 different models that they had the 4-o, the four, a mini, the four or 4.5. And some of them are reasoning models. Some of them are more general use models. And you kind of had to pick and choose which one you wanted to do based on your task. And that wasn't always clear. Now GPT-5 is replacing everything. Uh, everything is deprecated. Yeah. Just you'd be D five and a couple of versions with like a GPT-5 Mini and Nano and Plus, but it's all GPT-5. And the idea is that you give it your query and it has a filter or router that will. look at your query and automatically determine does this need a reasoning model and something more advanced or does this just need a regular normal transformer processing? Sure, sure, sure. mean, pros and cons of that because it's probably in their best interest to route you to the least compute intensive process. so maybe it's annoying that you lose was just like a giant cost cutting thing for them. Possibly. mean, the API model calling the API is also really cheap, like on par with Gemini Pro. 2.5 and way less than Anthropic. I don't know about you, but I was expecting GPT-5 announcement to be like, we have reached human intelligence. was of what was hyped as a bit, or people thought it might be. But I mean, yeah, I was trying to refresh yesterday because they said it rolled out to everyone. And I didn't get the update yet, but I have it now. I haven't really messed around with it. But in the demos, a lot of it was just like, yeah, it seems better. There wasn't anything that seemed like, oh, you could do something you couldn't do before. Two things, I think, is going on here. Again, I'm speculating and just for the entertainment of our viewers, go along with me here. So the one thing, and I kind of texted you about it, I was like, could Meta's poaching have actually worked and decimated the talent pool at OpenAI? I mean, I think, feel like this was probably way in the works before the poaching started. Sure. So. But you remember, they had to like shut down the company for a while because the poaching was getting so intense. Was it from the poaching? it shut down from the poaching and not from like the whole board drama? Fiasco, like what was that last year? Some of that as well. And the other thing I'll say, and I think I'm hitting a vein here, could AGI and ASI be like nuclear fusion? Like it's always two years away. This dream, but like never quite achieve it yet. And could that actually then? caused the AI bubble to burst. Yeah, right. It's like, well, is it because we need more compute? need more. Yeah, it's like scale will solve everything. need But clearly, GPT-5 has had the scale and the money, right? And it's still marginally better. Maybe it hasn't hit human intelligence yet. It's probably better. His comparison was like, what, GPT-3 was like a smart high school kid, GPT-4, smart college. person, GPT-5 is like an expert in everything. But he didn't call it AGI because he says like, I guess one of his benchmarks is that the model keeps learning itself and this doesn't. And the Silicon Valley folks are always looking at it through how their processes can be improved and. So they're looking at it through software development, right? Because that's what that industry up there is. And yeah, I think in terms of building code, and we just talked a little bit about vibe coding, sure, maybe GPT-3 is the high school kid and GPT-5 is the expert. And one of the big features I pushed on GPT-5 or demoed was like, oh, look, it has a whole coding thing. I feel like it's that more just they're catching up to Claude, not like, this is light years ahead of. They're literally forgetting that, you know, we live our life outside of co-development like everybody else outside of Silicon Valley. Like for example, I love cooking. love recipes. I look at Michelin star folks and go to nice restaurants. Like that's my thing. That's my jam. I'm a foodie. When I asked ChatGPT for a basic churro recipe, can't do it. Like the gets the ratios wrong. It misses. There's no eggs in the recipe. Like, I'd curious. Yeah. We try that with five. Cause I mean, they're like, you know, in their own benchmarks, it's like what's hallucination? What's this? Yeah. It'll never say like, this might not make you the best churros, but go for it. It'll be like for the best churros ever. But nothing, nothing that was like even, you know, a big functional use. Like, uh, it's not out of Google yet, but they demoed it at the Google IO where it was like, Oh, you can take your phone and kind of add in real time showcase. Like, Hey, what's that thing? And it was like having a conversation and looking through the camera. So like nothing like that. No update to the image model, which like, think, you know, chat to be the image is great, but nothing featured there. So like nothing changed with that. You know, I'm sure it'll just be like a better experience if you're free user, a better experience. Cause they're rolling this out to free users on the same day. I mean, honestly, probably the biggest thing that got the most buzz was the, um, absolutely insane graph that they displayed. Oh my God. What did that even mean? That might've been a vibe graph, but basically showing the benchmarks. But, um, if you actually look at the numbers, the number on open AI, oh, three is higher than GPT-5 without thinking. So yeah. And it's saying. graph. I didn't understand either of the axes or the color or where the color. mean, the color and the shape are designed to be like, look, it's greater, but there is absolutely zero scale to this because 69.1 and 30.8 are literally the same. And then 69.1 is also obviously way higher than 52.8. think that graph was generated by ChatGPT. That's why it's lacking a little bit. vibe grass. When people say, you know, a producer on a movie will be replaced by AI, I really now come back to like, where are we at? Who said that? My neighbor, the guy at that. Someone on Twitter said, rest in peace, Hollywood. So we must all be gone. Yeah, it's like, or, okay, let's like lawyers, a lot of people say entry level lawyers will be replaced by financial analysts. No, they won't. Not anytime soon. Not at this rate. No. Yeah. There was a joke after it was like, oh, like lawyers and accountants like breathing a sigh of relief after seeing the GPT-5. Absolutely. Yeah. I don't. Yeah. Nothing. Sorry, Sam Altman. But if you want to be on the show, give us a talk about it. Yeah. The other thing they did ask though, uh, earlier in the week is, uh, the release in open source model GPT. love that. Oh, I says, um, I'm all for open source models. The first open source model they've released since GPT. This is back, what, five, six years ago? while ago. Free AI. And also, you know, it's in their name. So going back to like their ethos of their name. Yeah, you can call them closed AI. do remember that. So yeah, they released the two versions of a model. I mean, they're less powerful than their top line models. Yeah. But the big update here is that it's really performant, even more so than DeepSeek. So you have the light-weightness and the performance. combined into an open source model. think that can be really interesting for putting an LLM on device. So it's doing local inference. for example, if you have a car and you want to talk to the car, Like Mercedes has their MBUX system, like a lot of virtual assistants in car. This could be a model that goes in there instead of whatever they have. Right, right. Or if you're want to run something locally or run it on local data and make sure doesn't go to the cloud. That'd be the other. Like if you're a military contractor, you can't have any internet access to anything. Or if it's like a tank in the battlefield. if you're a Starlink, it doesn't work in it. You need to have local reliability. updates from there. I'm more excited about the open source model than GPT-5 in this case. Yeah, I'm curious to how that plays out. people do with an open source, open AI model. Yeah, and the other thing with open AI versus Chinese model like DeepSeek is that companies will be more tuned to adopt that because it's legally in a more clear situation. Yeah, it's like an American company that you can go after if you have an issue. Exactly. All right. Other update in the world of media. 11 Labs. very powerful, popular voice over generator, text voice generator now launched Eleven Music, awesome music generator. I love music models. I love music models. I think it's great. It was a lot of fun. This one is on par or even better than Suno. I messed with it a bit. the other, things that it really excels at, um, is it has a really excellent interface. And it'll generate the music, but it'll also generate it in sections and you can drag or modify each section of the music or kind of adjust the timing so that the beat changes at a different spot. Intro outro chorus. So for generating, yeah, just kind of the background beat music and having a little bit more control over when things pick up or come down. Excellent interface so far. That's how they built it. Yeah. End of the day, how you interface with AI is ultimately what breaks or makes a model, right? Yeah. It's like, how can you harness 99 % of that power of that model? If you have a clean enough interface, you're doing it. But if it's like a text prompt or something, you're not. Also, other thing is they did work with some of the music labels. And so they're saying that the stuff it generates is legally cleared, commercially cleared. That's a big change too from what we've, where that it's been a bit of a gray area with some of the other music. mean, really puts into danger a lot of the artists that make their living from stock music. Yeah. The background beat kind of stuff. And the licensing that goes into it, they're getting passive revenue from like anybody that downloads it. I mean, if you've ever made a YouTube video, you're probably going to some stock footage website, grabbing some music tracks, putting it on your YouTube video. Yeah. Yeah. We have a bunch of. stock music subscriptions and stuff. But I mean, I've also spent hours looking through stock music sites. And it's just like, now what I'm looking for, now what I'm looking for. I love artlist.io. That's one of the better ones. Is that OK? I've used Art Grid for their footage stuff. Soundstripe is one I use a lot. They have good stuff. Yeah. Envato is OK. Envato's hit or miss. It's hard to find stuff because you just have so much stuff from some good stuff, some not so good stuff. Yeah. I've seen there are some other companies that launched that had deals with record labels where they had like quote real songs, but that you could license at like a more affordable rate for like a YouTube video. Oh, so if it's like a Coldplay track or something. Yeah, that they had a deal. that was always an issue before because it was like they weren't really in tune with YouTube creators. And it's like, oh, you want to license it? Yeah, it'll be 50 grand. Because it's like that standard rate for commercials or TV. And it's like, I just want to put it on YouTube, not get demonetized. And it's like now they have more. They've wisen up and made it. more affordable licenses and And YouTube is so good at sniffing out licensed Oh, their content detectors. Annoyingly excellent. Yes. Yes. The downside of that is they don't really have a good way to filter out if you legitimately used stuff under fair use. Right. So sometimes their content detector can be a bit too aggressive. Yeah. So I mean, out to 11 Labs for dropping something that the market generally needs. I mean, lot of the AI stuff that we look at on social media, know, TikTok, Instagram, LinkedIn, whatever. is missing the sound component. Yeah. In general. Yeah. Yeah. It's like, if you have you ever done actual content production, that's like half the battle. And that's yeah. Like bad sound will ruin anything. A video. Like mediocre video. You can get away with more than really crappy sound. Yeah. If you combine this with that Adobe tool where you can like the sound effect, I mean, there's some other models too, where it's like, can give it the video and it will analyze it and try to create the sound that matches the video, like the Foley and stuff. And then you're more aware of this world than I am, but just taking like vocals recorded on like a crappy source, like a phone, and then turning it into like master audio from a microphone. mean, Adobe Podcast, their audio fixer is, uh, is really good. Yeah. But for that's more for, you know, this kind of stuff where we're talking, we use that when we had that bizarre issue on, um, At the AI on the Lot. don't have issues. Not in here, but yeah, AI on the Lot when the mics are picking up some random, um, radio station and, uh, Adobe podcast was able to clean that out and, make the voices sound good and get the background audio, but it's not for singing. I'm sure there are other models that I'm or tools that I'm less familiar with because I don't deal with like music as much that. Yeah, we're not on the audio side. take a singing voice and make it sound like off of a crappy microphone and make it sound better. Well, it seems like Eleven Music is able to generate vocals. It generates vocals and yeah, that's hard. And then it takes, oh wow, it could do lyrics. Yeah, I mean, the one thing I liked about Suno is Suno had a checkbox option would be like instrumental or lyrics. And usually, I mean, the stuff I'm doing, I'm like, I don't know, this one. background instrumental music. 11 Labs, think I only mess with a little bit, but I would type out a prompt and be like, I just want a background electric beat. And it would add lyrics into it. So I think in the prompt, have to specify no lyrics. Or you have to put it in the prompt. There wasn't a checkbox to be like, just make it instrumental. Whereas Suino had a little bit better interface for that. I'm sure they'll get the feedback and get it right on the second release. All right, next one, Houdini. Houdini dropped, Houdini 21. And for those of you in visual effects, filmmaking, you guys all know that Houdini is the king of Sims. I mean, there is no other package that is doing, that is recreating reality, quite frankly. So like fluids, yeah, fire, smoke, water, destruction, like, you know, the collapse of a building or a car colliding with something. Even down to Joey, like muscle Sims and yeah, I'm seeing like a muscle of a dog. Yeah. So demoing right now about five to 10 years ago, this stuff was so computationally expensive, but GPUs just couldn't handle it. Right. And of course everything you're looking at in this demo here that we're going to show you, this is not real time. We know this, but that's okay because this is the highest level of quality you can achieve today. And control and control, including AI solutions, including unreal, including Maya like there is no other package more capable than Houdini. And the fact that they've upped themselves this high only assures like, we gotta stay grounded in reality. Like we talk about AI tool sets all the time, capabilities all the time. I don't get up cookies into dust. That's an end to the whipped cream on top. Yeah. Like when that comes in, it's the consistency of whipped cream and has the, like the material feedback. Yeah. It's so accurate. So look, if you're making a film at the highest level, you're using Houdini. You're probably not even going to touch AI because it's not even close. And you're not going to touch Unreal because you make a lot of sacrifices when you go to real time. So the simulations are just not as accurate and things like that. it's really good to see that we talk about a lot of AI tools, but end of the day, the highest level of accuracy and detail to reality is still achieved by something like Houdini. So what are some of new updates in the version 21? This was like a version 21 sneak peek video that they teased. I think they're going to announce it or do a keynote SIGGRAPH. Yeah, I'm not, I was never a Houdini artist. I knew some of them and they're all a little quirky because the Houdini is like the perf, you have to have the perfect 50-50 split in your brain, half technical, half creative. Whereas with like a software package like Maya, you could be fairly technical and get away with using it. Packages like Unreal, could be quite creative and because it's so easy, you can get away with it. But Houdini is quite challenging because it has a note-based thing and I also need to script in it. have to recall. So I don't run into too many Houdini artists, especially at like a tier one level. I remember working with somebody at my last job who was a world renowned Houdini artist. He was so valued in fact that Houdini just hired him. So there's not that many of them in the world. I think that's their only caveat is as capable as it is, it's really difficult to. Yeah, very specialized. Yeah. Yeah. So do you know what specifics it gives? Yeah. So I think it looks like the muscle sim stuff has up. Everything seems iterative. There's nothing new that they've never had before. The sims look better and has probably a better algorithm that they're running on. Right. I'm guessing maybe you could process or run stuff on. more local machines, you to run it faster. Everything is gonna be local. So typically how Houdini artists work is they'll set something up and then we'll submit the sim and then the sim will take a few hours to render and then they'll come back and look at the results. Okay. Yeah. And it only helps that the AI race is making us. get better GPUs out of Nvidia and AMD. That is true. So that's all going into you. You're to do stuff like this. Yes. Yeah. It's all net positive. All right. Awesome. Pretty cool. All right. Next update is from Lightcraft. So Lightcraft makes Jetset, which we've covered before. We've before. And I've done a bunch of videos with them. Basically, like a very accessible virtual production. You could use your phone as a tracker, put it on top of a cinema camera, and then kind of shoot, record all the camera data, and then later on go. link it up to an unreal scene or a blender scene and it makes very accessible virtual. Yeah. Lightcraft is also operating in a space with practically no competitors. Like they're a unique offering and the CEO of state art. there's either they're either like dedicated apps or it's like you have to shoot everything on the phone and it'll track, it's like you're shooting on your phone or, you know, or just do everything in unreal. then unreal doesn't have like physical tangibility into actual, like, you know, hard goods, if you will. Whereas Lightcraft integrates back to Unreal or Unity and then you have the ability to then use a phone to do a lot of that task. Exactly. Yeah. And so now they're sort of expanding their tool ecosystem to sort of be more of a one-stop shop for indie filmmakers and other productions. So they're launching a couple of web platform tools next year. So one is Spark Shot, which is... It's a browser-based tool. Yeah. You in the 3D scans, you can kind of bring in your models and like collaborate with your team and sort of... previs your film on a web browser, collaborative web browser. This is really powerful because as you know, typically USD files for filmmaking is very heavy. You're talking gigabytes, you know, and then to load it into a browser and have it accurately sort of spool it up and play it back is quite powerful. My guess is they're spinning up a virtual machine on AWS somewhere and then tying it into your browser session. Like a stream that, um, pixel stream streaming thing. Perhaps it's that, or perhaps it's, uh, based off of HTML 6 maybe. Oh, okay. Yeah. So there are some technologies, but the fact that his browser based means you can run it on anything. An iPad. Yeah. Yeah. And a lot of it, cause I think it's like cool too, that they're expanding it. Cause the demos have been like, Oh, if you want to do kind of a like low budget virtual production shoot. But you know, lot of other use cases too, or it's like, okay, maybe you're not doing virtual production, but maybe you're doing, you know, just previs for like a shoot on a camera. But it's like, have a scan of your location and you want to kind of bring in a virtual camera and like map out in pre-plan in your location that you're to shoot at. You could do it in a virtual space and, know, make sure everything fits, make sure your measurements happen. and do it in a nice collaborative web browser. Yeah, mean, just savings alone from not reshooting and not having like, you know, a second round of shooting is going to pay for like, yeah, especially you're on a lower budget and you don't have, you there is no reshoots. You only have one shot to do this, like literally. Exactly. Other tools, Spark Live, which seems like more of a kind of collaborative zoom-ish thing. So you could all kind of hop on the same collaborative web meeting. And I'm guessing like Shotplan or Do. work in the 3D space and then Spark Atlas, is challenges the artistic centric database built on the open source super based SQL system. Yeah. So that's like a, for the indie filmmaker, like a shot grid for your production. To track your assets and stuff like that. That'd be great. Cause also a nice, more affordable, accessible shock grid. would be, would be real nice. Welcome for the smaller budget community. mean, if you're working in light craft. then as any filmmaker, you're already got thousands of assets that you have to manage. then you need an management in the AI space, and I mean, they did kind of market this as like blending, know, filmmaking and AI, and it's sort of like a one-stop shop. But you know, the projects we're working on, I've been messing with the Shotbuddy that we talked about a while ago. just any tools to like track AI, the images that go to the shots and then the versions of the shots and then you're reprocessing the shots. It's still, know, there's so much overlap in the VFX pipeline that's already established. Uh, able to twice. But being able to have a tool to track and manage this stuff, especially on bigger projects, like gets really, really important. And then Spark Forge, which, uh, it seems like it some overlap, but streamlines the processes by combining asset storage, metadata management, and batch processing into what like craft calls a high speed shot factory. Yeah. So that's their batch processing tool. That's interesting. Maybe it ties into some sort of cloud render. Using possibly. Yeah. But I think it also ties into, cause they have a, can't think of the name of their app, but basically like if you are shooting virtual production with Jetset, you can load in your unreal or your blender scene on your phone and you get like a rough previs, but then you still have to take the data from the phone, take your recordings that you shot on the camera and process it to marry it later on a computer. And so this might also automate or streamline that process. So it's like, Oh, just hate, take all our takes. take all our takes and then just marry these. right. If you have the right naming convention, which I'm sure they help with, then it should just all line up. Hey Joey, how about this? Combining the Lightcraft spark with the Gemini world build. Dude. man. like, oh, instead of, yeah. Instead of like a Maya scene, you're now just generating on the fly. Yeah. I mean, I feel like that's gotta be the next Don't you know that Lightcraft CEO from your interview? Yeah. Eliot. Yeah. Yeah. You should shout out. Shout him out. Oh yeah. No, that's great. Yeah. We got video. got, we got like, have a series of with Eliot, like, like a four episode series. Yeah. Eliot's great. Eliot's also so much fun to talk to because he was like so passionate about this and he's so excited about this and so excited about making more accessible, like virtual production tools and AI tools and marrying the two together. Yes. And they also work with a people where you could like relight your shots afterwards, which also works really well if you are shooting on a green screen stage. in a virtual space, but you don't have the money to like have interactive lighting. You can post process it later. Yeah. I got to say like, uh, Beeble's relighting tool is probably the best relighting with AIF scenes yet. Yeah. Yeah. With no training, no training. Yeah. Right. You don't have to train your models. Give it an HDR map. Yeah. Netflix did a version of that, but their version it trains like you need to train the actors. You have to give it, you have to shoot the actors and give it that training data. Yeah. And then it'll relight, but yeah, people, you don't have to train on anything. Just give it a shot and it'll relight it and give you your Yeah, and then the last update from Bebo, we verified that they take now HDR maps. And yeah, I think they can generate them too now. I think Poon just said that it wanted So then that ties directly into VFX pipeline. And they do spit out, I want to say 16-bit EXR. Jeez. I want to say that was what they can export. Because remember, that was a comment on one of the videos. Yeah, those two words in AI typically don't go together. No. 16-bit EXR, AI world is 1080p, JPEG. I know. I might have some of the specifics wrong, but at a meetup, just saw him talk at someone asked the export option, he did say EXR, which was, yeah, think an early comment of like, it's not going to be in the real VFX pipeline until it does EXR. VFXR as well. It does EXR now, I believe. I'm 99 % sure that was what he said. Yeah, that's cool. Yeah. All right, other AI update, another image model from Alibaba, Quinn. Quinn. 20 billion parameters. 20B is no joke. Jeez. It says that it excels in text. Yeah. Accurate text, also both in English and Chinese. And it's open source. It's from Alibaba. Yeah. Typically Alibaba Tencent models are just grab them, use them. Yeah. I think one is Tencent, if I'm not mistaken, or Alibaba. One of those two. I think it's Tencent. Yeah. Yeah. And then you have, uh, Quinn and then you have C dream from a bite dance. Yes. Yeah. Yeah. Like these are really good, powerful open source models. Now is it ethical and is it, yeah. Like, it ready for Hollywood views? Probably not. Well, yeah. Based on which country, but you can't ignore what they're doing here. Right. It's like, if they get text, right, that's one of the hardest things to get right. So if they can do it with 20 billion. parameter training and some of the other stuff they're running under the hood, ethical models should be looking at that and trying to replicate that. Yeah, we'll see. All right, then Grok Imagine. Oh, Grok is back in the news. Grok is back. Did it do something bad? Well, mean, OK, first off, they released Grok Imagine, which was a new image and video generator, which would also let you make not suitable for work content. So topless models. Elon's all about, yeah, not holding back. this and the sexy chat bot that you can talk anime one? Yeah, the anime I think I saw it in Japan when I was there. So yeah, look I mean quality wise or whatever I've mixed things apparently generates really fast it seems kind of more like novelty than something that would go in professional use also there are issues like there was an article that someone was able to generate Uh, topless Taylor Swift new deep fakes using crock. Oh no. Imagine. So obviously like problematic stuff. Taylor's legal team is all over this one. sure. Yeah. And you see that thing that, um, LeBron James is legal team sued the creator that w it's making pregnant LeBron James videos. And they, sued, uh, they filed a cease and desist to stop publishing. Okay. I can see that coming after that. I would say that probably falls more into parody, but obviously like new deepfakes of any of any real person. Their permission is problematic. Yeah, that's like the early days of Napster when the record labels were suing Napster. And then we ended up with Spotify and Apple Music and YouTube Music. So you're saying we're going to have like a celebrity deep only authorized only fans? I we'll get over the novelty. But the fact that we can modify anybody's likeness is not going away. Yeah, I mean, I The yeah, what needs some, well, I mean, most models have safe or like, you know, do more thorough testing to make sure stuff like this doesn't happen. Yeah. Yeah. As Grok and xAI is bit more everybody should have a say in how their image is being created and altered. I, know, as a likeness and someone's making your likeness without your permission. there's defamation. And if your likeness is being used to generate money and make you look bad, you should have a say in that. But what I'm saying is like five years from now, we would have seen so many pregnant Lebron James that we would have just like, so what? Okay, what are you doing with that technology? Yeah, it's like, all right. That was so 2025, dude. Like what are we doing in 2030 with that technology? How are we changing the world? How are we making money? And how are we improving society? real life hacked LeBron James in your eye implant and you like think you're actually talking to real LeBron James? I don't know what the answer is or maybe it's LeBron James's likeness will never go away, right? Like it'll be a persistent thing, especially if he signs a Nike deal the same way we have Jordan clothing and Jordan shoes and all that stuff. Like it'll be a permanent fixture to pop culture, like the whole LeBron. I that or I mean, aren't there already some laws in the process of like, know, deep fake? banding laws or like crackdown on deepfake stuff. I mean, or it becomes like a law and then you have enforcement and then it's up to the companies where they face legal fines that they have to ensure that it doesn't like, that the tools can't make something that looks like a real person. Yeah, sure. Yeah, but yeah, I mean, you can't enforce it in China. Yeah, or if you build an open. source or you take a model and modify it. I mean, the person could be liable. I mean, they'll be curious to see what happens with the LeBron James model. Yeah. Yeah. I don't know. I mean, obviously problem. mean, look, a parody thing or pregnant LeBron James is like, OK, it's all we all know. It's kind of parody. But like a celebrity nude deep fake. Yeah. You know, more problematic. And then even on a a, you know, on a like local level on a. high school, know, making like making someone. Yeah. Some high school kid feel bad. And then that leads to making fake images of like someone, you know, for like cyber bullying. that's not there. Right. We need to have. my question is, what is Grok and Elon doing, you know, with this kind of like they shouldn't be trying to make money. I feel like there's no there's no money in this stuff. Maybe this is the other Westworld where it's like they make the the the crazy AI model so that the somehow the Tesla cars can drive better. I don't know how this one completely ties into it, but I don't know. Yeah. Well, it's out there. If you want to use grok imagine, tell us how it is in the comments. Yeah. Let us know if you've used it over any other model. Yep. All right. Quick one from a ComfyUI. They released sub graphs. So this is basically an easy way if you have like a really complicated comfy UI workflow, you can take like a piece of the workflow. put into its own graph, and then on your other workflow, can call up this little subgraph, run the process, and then come back to your big workflow. So like a very long requested feature, but a way to clean up and. Yeah, it's a way to collapse a portion or all of your workflow into one little thing and then tying into another workflow. So now you have super workflow. Also good too, because I'm guessing you could probably reuse this subgraph. So like if you have a process you're recurring, you could just take this one and plug it into other. workflows. Yeah, I would say the like Mr. comfy anonymous, like the one thing, one thing you could probably do is build a UI builder on top of comfy. So then if somebody wants to create a very simple UI with comfy UI under the hood, so you don't expose any of the nodes and the graphs to the user, but the developer has full access to it. Like a Right. I see what you're saying. Yeah. So right now you can do it with other things like Python, Pi UI, like that. You need something to host to process it. So I think there's probably third party companies because I know there's companies like Playbook sort of pivoted to this where they'll host, they'll run hosting for Comfy. And I feel like it probably be more of like up to them to build that easier interface. they wanted to, if you as a creator wanted to have a more front end facing user interface, cause you would need some cloud compute anytime someone went to it to run it to process that. Yeah, I would say it doesn't have to be Figma. Figma is like the jam right now in UX and UI development. Nothing is better than Figma. Take 1 % of what Figma is doing and apply it to Comfy. And you've got a winning formula where you're not only doing custom workflows that are hyper powerful, but then you're wrapping it up in a nice little box and giving it to a user. Next update, more quick roundup. Claude Opus 4.1 was released. So an improvement over four. I don't know what exactly improves. I haven't tested it out yet, but see a fancy graph. Let's go through the graph. is accurate. Looks accurate to scale. I mean, the one here focused on software engineering, but yeah, Claude still one of my favorite models to use for writing and a lot of other tasks. For sure. Google Labs, have their own. code writer called Jules. And it is now rolled out of beta. And if you're on the Google AI plan, I believe you could tap into it. And I've heard good things about it. I can't remember all these names, man. I know. I even know Jules existed. No, but I am Jules. Yeah, I'm curious if this is more for actual coders or if it's kind of a vibe coding tool. My guess is they needed it internally anyway. So they just built it. And now they rolled it out to everybody. Sure. Good point. Yeah. And the last quick update is Mid Journey, which had video and does pretty good video outputs. They now upped the video outputs to HD resolution and some of the demos. mean, some of these are just crazy detailed and look really good. Yeah. The video generation world, seems like are now more or less by default 1080p, right? That's becoming more more standard. Yeah. Eliot. They only do 1080. Right. And I think is the runway Aleph at 1080? I think it is. Yeah. Vio is. I think so. Veo 3. Yeah, Veo 3 is. Yeah, Seedance is not, I think, Luma 720 still. Right. But yeah, but they're coming more and more picking up and making 1080. Yeah. Standard. Yeah. I would imagine in a year or two, now you're looking at 4K, hopefully. Actually, I did see a tweet, because I think someone asked, Chris DeBall, like, when can we get 2K, 4K outputs? And you're like, soon. Yeah. So I'll say it to Eliots defense. Like he delivers. He's the anti opposite of Elon Musk. Like self-driving coming in three months. It's been three years. Yeah. Eliot brings it. Yeah. You also do the tweet out another tease of interactive, like world model from runway to lock and create like that thing we saw. Like, I think he said end of year. Okay. We'll see if that comes. The one where you can just kind of talk. I if it was that one, but something that was, think, cause you know, he sort of. responding to when Genie 3 videos were out of like something in that realm coming soon from Runway. Yeah, they're really, all the AI companies are really quick to dismiss and one up the other one. Yeah, was like, don't worry, we're talking, we're working on this too. got this coming out. You think that's a big deal? Wait till you see this. We should do this with other podcasters, shouldn't we? What? You thought that episode was great. All right. And that was pretty much the round up today. Let us know if we missed anything. yeah, links for everything we talk about as usual at denoistpodcast.com. Thank you for the YouTube comments. Somebody just commented just to feed the algorithms. Yeah, thank you for feeding the algorithm. Yeah, thank you for that. We could use a couple of five star reviews on Apple Podcasts and Spotify if you haven't done so. You know, we're still in a double digit number. Like it makes a big difference. So please help us out. All right. Thanks, everyone. We'll catch you next week.