Nano Banana: We Found the Best Tips & Use Cases

Nano Banana (Google's Gemini Flash 2.5) is transforming image generation with capabilities that put it ahead of the competition. In this episode, Joey and Addy explore impressive use cases, best practices, and technical tips for this powerful tool. Plus, we analyze Mickmumpitz's innovative Comfy UI workflow for VFX that seamlessly blends live action with AI generation without green screens. Finally, we examine how AI is reshaping the music industry with synthetic bands, AI K-pop idols, and even record deals for creators with zero musical background.

--

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

0:00

All right. Welcome back to Denoised. We got a couple stories to jump into today. A lot of Nano Banana stuff. Nano Banana lives. It does. It thrives. It thrives. It's, it's growing. It's mutating. So yeah, turning into a jumbo banana. All right, let's get into it. Okay, so first big story, Nano Banana. Not to too our own horn, but we were, we were like pretty ahead of this game. Yeah. Before everyone caught onto Nano Banana and, and Google. For a minute there, if you googled Nano Banana, our video came up first. It was in the very top results, and some people were messaged us like, Hey, I was trying to search Nana Banana, like your video was showing up top. So Denoised, you heard it here first. Here, you're on top of this stuff. That's right. But now that it's- More, more the reason to subscribe, like, follow. Exactly. But now it's out in the wild. Google confirmed that, you know it. Um, it was their model and the official name is, uh, Gemini Flash, 2.5 Flash 2.5. Flash. I'm glad they are still even. Externally calling it N banana. Yeah. It looks way cooler. Way cooler of a name. And all the thumbnails on YouTube and everywhere is like the banana. Yeah. People are generating a banana. Banana, yeah. Into bananas. Yeah. It's great. It's fun. It's a lot of fun. So, uh, it, it's already out. We're not gonna cover like, the details of it, but there have been a lot of really awesome use cases of it. And I kind of wanted to go through some, uh, tips, best practices, things I've seen online. Interesting use cases. Yeah. Uh, to kind of just showcase what I've been seeing, what you've been seeing. And uh, yeah. And then the more stuff I see from it, the more I'm like, damn, this is the most powerful image generation model to date. Yeah. This has been my go-to, uh, for a lot of people modification. I still have not used it as much for like generating stuff, but if of like, I need to modify an existing image. Yeah. Or something that generated or tweak something, this has been, this has been the go-to. Yeah. Like for image likeness, if you wanna replicate yourself or you know, a digital actor. This is the go-to if you want to. You know, throw a prop or a product into somebody's hands. This is the go-to. Uh, so let's get into some of these big use cases here before we jump to the other use cases. I think we should probably just showcase the desk, the test I did.'cause that, that, that makes the most sense. But, uh, yeah, I did a fun test where I just took a screenshot from this podcast at this exact table. Yes. And then I, uh, for one test, I grabbed an image of an old clamshell, uh, MacBook. They weren't even Mac iBook. Yeah. Before they were MacBooks and swapped out our. MacBooks with those, they really depended on finding a good source image. And surprisingly, there were not a lot of good source images of just the laptop open. Oh yeah. Was they're all like the 'cause they used to. Product, display them like this for whatever weird reason was a thing that Apples do. It was cool. Yeah. But that was like the bulk of the product photo. So I had to find a good product photo. Okay. I had to change the camera angle. They're phenomenal. I change the camera angle, so like I said, make an aerial shot. And this thing, the thing that's crazy to me about this is like your cup, your mug, this was like just a sliver of it was showing Yeah. Behind your laptop and it like nailed the complete mug in this aerial shot. Okay. You wanna know something really creepy. Yeah. That wood flooring is exactly the color of this wood flooring. How, how does it know that It is, and it's also funny because we're on a rug, but on top of the wood floor. Yeah. And like there's none of that in this shot. I think maybe it's just. Assumption based on the vibe of the room. Yeah. And, and it got like, as far as likeness goes, like the right side of my body or the left side of your body, it nailed. Yeah. And even our arm position Yeah. Matches pretty well. And then another one for fun, I just put us in front of a live studio audience. Uh, kind of warped. Yeah, my face a bit. Somebody commented that, yo, that's Will Smith territory. Can can we get into that later? We'll get into that one too. But also, what I thought this was funny with this one was it took the image from the shot and put it on TV displays in the audience. You see like. Oh wow. Put it, that's nuts. Yeah. Yeah. Did you ask for that? No, I just said put us in front of a like live audience. This was like, yeah, you know, I ran this a couple times. This was the best one. Did you see the reflections of the iBooks on the desk itself? Maybe it's a little exaggerated, but the fact that it's, you know, a good called, that's crazy too. Yeah. It has like the white reflection. Wild. Okay. So yeah. And then somebody else pointed out the wood grain on the top view shot is accurate. Because the table is in four pieces. Yeah. And the wood grain is in four pieces. Yeah. And yeah, the proportions look a little weird, but it, it, yeah. Figured out that this is in, in four sections, so good. Yeah. Other use cases that we've seen, Google had a good roundup of. Some of, uh, use cases that they've view found or some other people have found. So one is you can basically kind of like we covered in Veo 3, you can drop annotations on top of an image Yep. Of like what you want and it'll honor those changes into the new image. And Freepik actually just released a tool that makes this really easy. Um, you can't add, uh, photos to it, but it's an annotation tool, so you can take your image, click. Of like, Hey, change, you know, the headphones to a hat or like various precise notes and it honors it. Wow. Yeah. Wow. And Freepik's where I've been using everything too, just because they have the unlimited, unlimited credit thing, which really gets you. Yep. So that's been, that's been super handy. Yeah. Freepik's been great for just jumping across models really quickly. Yeah. It's great for testing. I mean, that's how I discovered Seedance. Right. And, and the other ones, 'cause they had'em and I'm like, well, I'll try it out. And then, uh, these are great. And cost effective. Yeah. What else? We got head shots change in angles. We, we covered that one. Yeah. So this is the one I sent, um, I'm looking at the third one here. So the architectural example. Okay. So taking just like a iPhone photo of a building, it could then generate like a arch, arch V, like a architectural visualization of the building. Oh yeah. The isometric view. Yeah. Yes. Yeah, that one's cool. Yeah, that one. A photo. Yeah. This photo that they showed is pretty. Not that great. Yeah, it's not that great. I, I, and I'm sure it's like filling in a lot of the missing details, but at the same time, this could be really useful for somebody that's doing city planning, urban planning, anything like that where they need to stick to a certain type of, you know, architecture or, you know, if you're doing rural building and media entertainment. If you're just trying to build a game or something, you need some assets. I mean, I was thinking this would just kind of be a fun, interesting way to animate.'cause you could, you know, create this isometric model of whatever you're taking a picture of, and then you could just run this through Veo 3 and say, Hey, make a 360 spin. Yeah. And, or some sort of other animation. And you've got like a model of a building of something you took a picture of. Yeah. And you could do all sorts of things with that in, in, in Veo 3. The other territory that, I don't think we have it here, but, uh, I'll, I'll just, um, saw, saw some online is. Image restoration. So now we're in like Topaz and McAfee area. Uh, right. But this one is just on another level. Like I have this shot too, where someone who restored the first, uh, photograph I ever taken. Isn't that crazy? Which was on a, uh, I believe it was a RO type, right? Yeah. And I didn't even, like, I didn't even, I. Put together. I always, I mean, I've seen this image before and I'd never really pieced together what Right. It actually was. And I'd seen this restored photo. Uh, I was like, oh, it's rooftops. Yeah. Saw I was just like, day I saw some of like Martin Luther King and stuff.'cause those were all black and white. Oh, okay. Like, whoa. I've seen to two have just like really destroyed, uh, looking, you know, torn and printed archival photos and just having them restored. I even saw one today, I think it was from someone that works at Freepik, and they gave it like the, it was, um. It was like a mugshot. Of like an old prisoner and then like a description of like the crime they did. And then it made the photo of the crime with the person. So it like read the context of the photo. Yeah. And then made like, kind of did a reenactment. Well, you know, of the crime. Um, I used to, well I, I still do, I watched the YouTube videos where they, uh, take historical figures and turn them into moving characters. Yeah. Yeah. Like they'll take Julius Caesar and now he's like, you know, drinking a glass of wine. Yeah. This will like level this up. Onto a whole different playing field. Completely. Completely, yeah. Yeah. I mean, it gets into that territory I was talking about in the Netflix AI episodes where it's like the line between the documentaries clarifying that it's like not real archival footage, right? I mean, you know, where it's sort of like, you know, you think of Ken Burns and it's like he did, he elevated, you know, took archival images and elevated them. It's like, no, there was not really sound there. So it's like in that realm where he is like elevating Yeah. Bringing history to life and it, it's, it's really moving and powerful because you already have historical context when you watch that movie or that TV show. So for you to like watch a lifelike reproduction of that person at the same time knowing the historical context really drives that home in your head, I think. Yeah. Yeah. And also just making, uh, recreate and stuff. Yep. That might have been shot traditionally, uh, you know, with those History Channel specials where it's like the interviews and the people talking, and then it's like the World War II reenactments and stuff. Yes. But, uh, uh, doing that with Gen AI, that's what, uh, gen, that company. Yeah. Uh, does, they're kind of specializing on lower budget documentary style recreations here. This was the, this was the mugshot with description. Yeah. Turned into Oh yeah. Reenactment of the crime. Yeah. The, the, the guy has such a distinct haircut and a look. Yeah, yeah. But even if it pulled the context of the. It, read this of what he did and then use, use that as a prompt. Like I think they just gave it this image. So that was what criminals used to look like back in the day. Very nice haircuts. Yes, very posh. Great fade. And then, yeah, one more. This was, uh, from, uh, Philipp Schmid Who works at DeepMind and he posted a list of best practices for Nano Banana. And I'll just run through a couple that I thought were useful. Uh, one be hyper-specific. The more detail you provide, the more control you have, right? In natural language, I'm guessing. Yeah. Provide context and intent. Explain the purpose of the image. For example, create a Lugo for a high-end minimalist skincare brand will yield better results than just create a Lugo. Right. Use semantic negative prompts. This was interesting. So instead of saying no cars, describe the desired. Seen positively, quote an example, deserted street with no signs of traffic. Sometimes those prompting tricks. I've seen this with other models too, where it's like sometimes they don't work well if you say, don't do this. Yeah, it works well if you say, I want this, which is the opposite of what you don't want. Right? I, I think Google. Inherently believes in as much positive prompting as possible, which negates negative prompting. And that's the whole reason why like Veo 3 doesn't really have negative prompting? I think possibly. Yeah. Yeah. It's like their ethos and their DNA, no negative, negative, no evil, no negativity, no negative. Uh, the last one, I think this one's most useful because you could, you can give it multiple images and it'll, you know. Generate the output, but people are having issues with the aspect ratio. Not being what they wanted. Yes. And I saw that. So apparently when editing Gemini 2.5 Flash. Okay. Nano Banana, uh, generally preserves the input images aspect ratio. If you upload multiple images with different aspect ratios, the model will adopt the aspect ratio of the last Yes. Image provided. Yeah. So whatever. If you want your 16x9 image Yeah. Upload that last. That's a hack to Yeah. Just do a little, make sure get the right in Photoshop. It's not that hard. Right. Well, no, I mean if you're like uploading, you have like, let's say we had our reference image here, but then I'm like, wanna give it an image of like a square photo I found of a hat to put on you? Yeah. If I uploaded our image first and then the hat image of a square, it's gonna give us a square output. Correct. So, so just take that hat in Photoshop, or No, I'm just saying No, I'm saying just upload. Upload the hat first and then upload this image first. Oh, okay. Last, because whatever we upload last will be Okay. Aspect ratio. Got it. So you don't have to do Photoshop or can you And save you time. Can't you just prompt and just say, give me a 16, nine output. Um, I don't know. I think people are still having issues for that. Okay. Upload it last Unusually manual upload. Just upload. Upload it last. Upload the last, yes. Okay. Yeah. You seen anything else with Comfy? With Nano Banana. Comfy with Nano Banana. Know why Comfy is on my mind. I mean, just, uh, there there's been a, like when Flux Kontext came out, there was a lot of outputs, uh, generating likeness with people, you know. Uh, whether it's themselves or somebody, or sometimes celebrities and then a banana, one ups, all of that, like what it do. I think what it does better than Flux Kontext is, uh, skin detail, wrinkles, pores, um, getting the subsurface scattering lighting. Correct. And, uh, so. I always talk about uncanny valley with digital humans. Like Nano Bananas, like one notch closer to closing that gap than I think previous AI models. Yeah. I found it where like it really follows prompts. Well, it does not modify other elements outside of what you wanna change. Yeah. Where with flux sometimes have that issue and moving camera angles in different positions Is, has been light years ahead of, of anything else. Yeah. Like I, I will say there is, uh, we cover this on the pod Enhancer.ai, which is the, uh, service that adds, uh, skin details, like, you know, whether you have, uh, freckles or whatnot. If something like Enhancer was built into Nano Banana or Flux Kontext, then that's like one less step you would have to take to make more lifelike humans. Okay. Or you know, it's probably just like you could build it into a workflow tool workflow or something like a Freepik or, or invoke or we, yeah. Yeah. Right. Where it's like, okay, you do your image and kind of get what you want, and then once you got. All your elements locked off. Yep. Then you enhance or modify. Yeah. A lot of the workflows that I'm seeing is like, it starts with Nano Banana, then it goes to enhancer if there's a person in it. And then it comes back to Veo 3, and then it goes into some other stuff. Like it'll go into Runway for VFX or whatnot. Right, right. Yeah. One thing, other interesting thing, uh, or useful tool, 'cause people have been like, Hey. This is great. Uh, but you know, I still live and work in Photoshop. How do I get this in Photoshop so I don't have to rely on, uh, you know, Firefly? Yeah. Oh, ouch. So someone, uh, Rob de Winter built a Nana Banana and Flux Kontext AI plug-in. Nice for Photoshop. Nice. And so, yeah, we got the GU link. Uh, it's 10 bucks or suggested 10 bucks. But yeah, this looks, you know, um, awesome to just bring these tools into the one for sure. The spot that you live and work in. Um, you know, especially something that's powerful with Photoshop. This is like. Takes to the next level.'cause you could just modify layers and you know, create elements that you want. Absolutely. But then still have the fine tune control that you get with Photoshop. This is the last mile thing. Like if somebody closes the last mile, that's where they make the money. Yeah. I mean, I imagine 'cause like Adobe has been rolling out like. Bring in other models. I think that's currently, if you use like the Firefly web app, you can like call up different models and the third party model things. Some of the Right, I would imagine it's probably only a matter of time before this is just built into Photoshop. Sure. But then they Good stop gap for now for 10 bucks. I was, I was thinking about Adobe the other day. I was like, okay. They were the first to drop a commercially safe model Firefly and uh, last couple of years their entire sort of strong suit was like. Everybody else's model is dubious. Ours is commercially safe, ready to go. And not that they're moving into third party models, like that claim is no longer valid. Well, I think the way they framed it was like, it's up to you and what your like legal. You know, uh, tolerances. And so it's more like we're gonna give you the option. So it's like, yes, if you want to create with Firefly Yeah. And keep it all commercially safe. Like choose Firefly, but like, but lesser quality. Yeah. Arguably. Well, yeah. Yeah. Well, yeah. I mean, I don't think I would say that, but yes. That's an ar not arguably, it's like what's more important to you, quality. Or legality Co. Right. So, uh, you know, they're like, we'll give you the option for the model and uh, you know, like if you're doing something, even if you're following Netflix's guidelines and you're doing previs stuff where you don't care what the issue with the model is, right. Uh, because you're just doing pre-vis, yeah. You load up Flux or Gemini or, um. Black Forest Labs or whatever. Okay. I'm gonna add one more. Okay. That's, that's the same model Adobe thing. Um, yeah, I follow some of the Adobe folks online on LinkedIn and stuff and um, lately I've been seeing a decline in the number of Firefly posts. Yeah. I don't really see, the only thing I see Firefly more so now is like Firefly as a. Like it's a web platform that is competing with the likes of like Freepik or Right. Flora or, uh, right. I mean, they just released a beta of like a, a mood board generator Right. Where you can call up different models. I have not seen Yeah. That much about the Firefly model itself. Yeah. I, I agree with that. Right. Yeah. So it's a, it's an interesting move and I, I think a. Company as successful as Adobe could definitely pull it off and figure out how to monetize everybody else's model. Uh, yeah. I will say maybe one argument for this third party script is being able to use your own API keys and 'cause I think that's also gets more important too. Cool. Especially with like client work. Yeah. Where it's like sometimes you just wanna have a dedicated API key.'cause that may be the, uh, client, you're doing client work and it's on their account so they're paying for it. Or you need to keep track of. Different uses for different projects. Yep. If you're doing this, if Adobe does roll this out and built into Photoshop, it's probably gonna use their credit system. I don't know how their credits convert to dollars. Yeah. Of like what the equivalent is, but you know, sometimes it might be a lot cheaper if you just load up your own API key and you pay the uh, direct fee. Yeah. To Google or, uh, fowl or whoever's hosting the model and then run it through there. Yeah. And that, that might be No, that, that's pretty cool. More cost effective. So that could be another argument for why even if Adobe does build this into Photoshop, a third party tool could come in handy. I have a billion dollar idea. Okay. If you're watching out there, you have a friend in Adobe, let them know this is for the improvement of the entire AI industry. Adobe has enough software development expertise, enough management expertise to really pull off a comfy UI competitor that's way better than Comfy. I know they have the expertise to do that. They could own the, would they professional market, the AI market in that sense? Right. So like, I'm trying to think of like, I can't think of anything that Adobe does that is node based though. Like in the world of like, you know, after effects versus fusion versus nuke. Everything is Fusion and NU node based after effects. They've not done node based, like they've not done, I can't think of any other software that has gone the node based route. Yeah, I mean they could figure it out. Right? I'm not saying they can't figure it out. I'm saying does it fit with Adobe's? But like now that they're backing away from having being a model company to now more of a integrator company, if you keep backing up. You could become a, a developer company. So like the same way Microsoft has GitHub and copilot Like all the tools that's, they're making the picks and shovels for the gold rush. Right, right. Adobe should be in that game too. I don't just agree with that. Yeah. I don't know if they would, I don't know if I see them building like a, something as. Going after like the comfy market of like the highly technical Yeah. Let me like build out my, you know, super, super custom workflow. Well, maybe it's like the technicality could be dialed up and down in the same way. In Unreal Engine, if you don't want to use any blueprints, you don't have to. Right, right. But you can totally open that up and give technical, you can even do c plus bus coating in Unreal. Yeah, right. I've heard of that. So maybe there's like a dial for how. Professional you want to get, but I, I think it's like a segment of the market that really has no clear winner yet. And yeah, we talked about comfy a lot. They, they're not monetizing, they're not really, I would see Adobe maybe going more after, like doing more of like a flora. Kind of thing. Yeah. Right. So that's, that's like the stepping up and down the complexity chain, right? Like at a intermediate level, it's flora, but then if you really want to build your own custom nodes and open up APIs, you just peel back the layer and get down to a comfy complexity level. Yeah, I mean, I'm also just, I've, I'm trying to think, like, I've never seen Adobe Open have any products that are like that open. Like if you, if I wanted to adapt my fal.ai. APIs and to nodes. Yeah. I don't see Adobe letting me do that because they would want me to like run through their credit system. Yeah. To, I don't know. Yeah. I mean, I was just thinking because of the Photoshop plug-in that you mentioned, it's like, wait a minute. They could have plug-ins for everything, every model, like they can own that whole, I think Adobe's target is the creator, the editor, the designer that wants to do the stuff and like wants to click buttons. And the thing works. Yeah. And not get into the nitty gritty of like, let me build out my, I agree. Okay. Yes. My node based workflow. Yeah. Right. Like Adobe's been more like, it should be a tool and a button that works behind the scenes and does the thing that you want. They're creator. You don't know how the magic happens. And user focused, but not, yeah. They're not Microsoft. They're not Google. They're not Perplexity. I remember seeing a post that was a good argument where it's like. Why did Adobe not turn Dreamweaver into like the web flow or something else? Like to cat to keep up with web development at the time? Yeah. Yeah. I don't know. Yeah. Yeah. Okay. I remember, I remember Dreamweaver. That was my first web builder. But look, you don't have to be Adobe. I'm just saying there's still room for, uh, a company to move into the AI developer space and really own that. Yeah. In the same way, like Google totally dominated the browser game and now Chrome is the basis for everything. Yeah. If you're gonna do any type of web developer, well, I mean. I mean, I know we have talked about this of like, how does Comfy UI make money? But like, even if you're not an active, comfy user, there are a lot of web portals and platforms where Comfy is the backend for sure. Like, like a lot of stuff that you're uploading images and it's just running through a comfy pipeline. Yeah. So it's there. I don't know what their future is because I don't, as we've talked about, we don't know how they make money. Yeah. Hey, uh, comfy Anonymous, if you want to come on the show. Mr. Mr. Mr. Uh, comfy Bitcoin guy. It's the same Devo, the same crater. No, this is a real person. Yeah, I, I sent you a video of them. There's a, he's not a secret person's name. What's, yeah, whats the Bitcoin guy's names? Satoshi. Satoshi. Yeah. Yeah. All right, we're back and rambling. We've got enough Nano Banana. All right. If you found any other use cases though, let us know in the comments.'cause um, there's just new stuff hopping up all the time and I'm messing with some other stuff too. I'm trying to blend. Banana Banana with Mary from Moonvalley, uh, for some workflows. So, um, I'm gonna, I'm gonna post that soon, but Nice. I found some interesting use cases there. It's just so good because it doesn't modify, it doesn't, it can, it'll leave the rest of the image alone and it'll like just modify specific things. The end painting is sharp. Yeah. Yeah. All right. Mickmumpitz. Mickmumpitz, our favorite. Okay. So McBuckets has a new tutorial, MCM we've talked about in the past. He's posted. Some of the most impressive, complicated comfy UI workflows. Yeah. Uh, that we've ever seen. And, uh, shout out to him for when I first started on Comfy UI, I downloaded some of his workflows and man, I mean, such a game changer work of art. Yes. Chef's Kiss. Thank you make for your contributions to the AI world. So he has a new, uh, tutorial app, but I wanted to run through this one specifically because. Usually a lot of this stuff has kind of been more in the uh, animation space. Yes. But this one is VFX with live action and it was one of the most sort of impressive workflows and blending live action with AI generation. Yep. And it's something we've talked about a lot of, like how these are gonna be the future and like blending these two elements. Yeah. And this is a really good workflow and so I wanted to kind of. Yeah, think, and it, it does mimic traditional VFX in the sense that things are layer based and you know, you're modifying, for example, the background without affecting the actor in the foreground. And things like that. Yeah. So the. The gist of the workflow is basically shot everything, live action. And would modify either just some elements, like modify him in the existing world that they shot, or completely change the world and keep him, and basically he was like, it was sort of like acting with a green screen, but there was just no green screen. Right. Uh, it was just in the real world and we just needed to get the positioning right. Everything else we could change with ai. Yeah. And he's doing essentially match, move without any camera data, any tracking information. Yeah. Which actually I'll talk about in a sec, but that led me to think like, oh, this would've been, this actually would be a really good use case for something like Jetset to give track that. The camera position. You should connect the two dots. You should reach out to me. I, yeah, I was gonna send Elliot this link to, he is probably seen it already, but I was gonna be like, Hey, this would be a really good use case for, for Jetset. Right. But yeah, so basically he shot everything, runs it through, uh, comfy, which he built this really complicated, elaborate, comfy workflow that identifies the subject. Isolates the subject him. Um, basically for the workflow to work, you need a, you need your original video, you need a rough cutout of the subject. He's like, it doesn't have to be great because the AI will just generate. The rest of the elements, right? Like, so it doesn't have to be a super clean mask of the person, but you have to have your person masked out. You have to have your, um, uh, black, white, uh, mat, uh, death mat. Uh, I don't even know how that'd be a depth. Uh, yeah. Is it death? Uh, just a ma like a mask to mask out? Oh, no, that's just a mask. Yeah. Just it was a black and white. Black and white. It wasn't a, yeah. Yeah. And then camera tracking data. If you want. And so we had a workflow too, to like track. Points and data in the scene, he is like, or you could just track it in another app and just pull in your, uh, tracking point. Um. Coordinate text file data. I, that's where I was thinking, oh, jet Lightcraft Jetset would make right a lot of sense because that was like an extra step and you had to like track and maybe possibly like bounce around a bit. How did you bring in camera tracking data into Comfy? It literally text like, oh, gotcha. Like a JSON file or something. Let me show you. Uh, it looks like this and I. Don't know enough about tracking, like coordinates, like just XY coordinates, like I guess this is JSON Yeah. That, that is tracking information. But you know, it's so arbitrary. There's no timestamp or anything. So, uh, I guess it doesn't matter. He covered it in a previous video. He is like, we covered this more in like our tutorial, but it was, um, some way that it works. Yeah. However the data's spit out. It just kinda schmos s interview showed it worked. Um, I, yeah, I mean he had a, he offered a few different techniques for tracking the camera and getting the data in a comfy, there wa I mean there was also a node that, 'cause this was only if you use the external tool, there's a node for Comfy, uh, co tracker. Point tracking, which will track, run your footage and, and, and, uh, it'll, it'll track the footage. Okay. So there, there's a preexisting node he's using there options, right? Yeah. To track, to bring in camera tracking data into a comfy. Cool. I didn't know that. I mean, comfy has all these nodes. I'm not aware people, I mean, where, that's the thing, that's the other advantage of the open source. Yeah. Platform is like people can, we just build, you could just modify it and build and, and, and customize it to do whatever you want. Right? Okay. So you got all this data and then. There are options to, basically you could generate a new scene, a background scene, and kind of give it a first frame. You can give it a text prompt of what you wanna change, and then basically it is modifying everything else that was not masked out. It's everything in the black and the. It would be flip mask. Yeah. It doesn't matter. Mask the non masked side of it. Right. Whatever you cut out from your subject that was not there, you can generate. So you can give it an image to help, you know, feed it or you can just do a text prompt. Right. And then it is generating the background using uh, Juan 2.1. Yeah, I'm sure because he's been working on this for a while. I'd be curious how this improved this Juan 2.2. Probably better. I'm sure a lot better. Yeah. And it's. Generates the background and then blends all of the edges with the character. Sometimes it gets like, uh, he set, he like shrinks himself and he sits on a couch. Yeah. And like the pressure from the couch from his hand, even though he's like literally sitting on like a box, like the, there's a depression of his hand on the couch. Yeah. And like a little bit like you see the kind of couch he like move. It's like aware of that Yeah. Of what is happening. Yeah. With the subject matter. So it it, for those of you who are following along the video, it's essentially like a Honey I Shrunk the Kids type of movie production. Yeah. Where Mick is, uh, tiny into uh, a world that's normal sized. Yeah. Here, lemme I'll. Run the, uh, final movie, which he has like before and after. He has like the original clip and the right and the final shots. I think in summary, basically what he built is like a Runway Aleph that you could run with comfy, way more precise. But the thing with this is it's like he is masking himself out. So like the stuff he filmed of himself right. Is not changing. Because the issues I've had with Runway is like you'll say, oh, modify this, modify that, and sometimes it'll change this. Yeah. But then it'll also change how the person looks. Yes. And so you always kind of get this like hit or miss. Yeah. This gives you that precise control of like, yeah, keep me, keep the actor the same, do not change them. Then change everything else. Yeah. This is more of a pipeline for an entire film. Where you can throw different shots and each shot would have a different requirement versus Runway where it's like a shot by shot generation. Yeah. Hit hit. Hit or miss. Hit or miss. Hit. Hit or miss. Yeah. These are wild too. Look, because this one, it's like literally he just needed to, like the whole city street is fake. Running. Also phenomenal job with high speeded motion 'cause he's running the camera's shaking and jittering. Yeah. And the generations and everything are still like sharp and work.'cause even like a lot of times you get issues with like high speed movement and like with, uh, yeah. Frame rate's always an issue with video generation. Yeah. But even just like a lot of motion blur in the shot with something, like if you try to give it to Runway, like it doesn't always do the best with like, uh, you know, with pans or super shaky camera For sure. And this stuff looks great and it's like he just needs to shoot. He's shooting down the street and he just needs to shoot. Yeah. The bus scene is, uh, he just needs to get the angle right of himself and like everything else doesn't matter. Yeah. The, so the bus scene was really interesting how he pulled it off, and it's the same way we pull it off in CG Splash. He modeled the bus with the people in the bus, in blender. And then did a camera move, uh, you know, very low angle. Very small. And then pulled, oh, there's the couch. Pulled that into. Comfy has a depth map and then comfy just generated the people and the bus based on that depth map. Yeah. It's really unique. Yeah. Oh, right now he becomes a giant in the city. So yeah. I thought this was Oh, oh God. Also, you know, he is running this entirely on his computer. So this is all done right. With Juan and, and other, uh, open source model. So this is all done locally? One person? Yeah. Or I think he's got maybe two people. Okay. You do the camera person and I think yet someone else help. Help him create the workflows. But this is such a great workflow. Yeah. That I thought was worth noting because yeah, it really blends a lot of the stuff. We've talked about this now when you say AI filmmaker, this is what I'm thinking. This is what you're thinking. Yeah. This is a proper film, a filmmaker, and you're just using like that's why the whole can just call yourself a filmmaker when you're doing like. Veo 3 text to video, but he is not, he's a filmmaker and he is just augmenting the VFX work with like Right. Additional AI stuff. So, yeah. So he's a filmmaker, right? That's what I'm saying. Oh, not an AI filmmaker. Well, I'm saying that right. I'm saying saying the AI filmmaker I feel like is, is everything is generated. And it's synthetic. It's like offputting and then the audio's from 11 labs and the videos from this. And it's like for now, yeah. I mean, sure. I'm sure that'll get better, but, and maybe eventually it'll just be filmmaker. It, it'll be a while. I, I think somebody like Mick will get there way faster than. AI filmmaker as well, or does he or does he even, is that what he is even going for? Or is it always gonna gonna be like a blend? He's gonna call him a hybrid filmmaker in today's, I think he's just filmmaker. You're, you're like a year ahead, two years ahead of where he made Yeah. He made a short film. Yeah. That, you know. Okay. You could have done this a while with, how'd you even do this? Pull off normally after effects and you give to build a virtual environments? All of that. All of the above. Yeah. Right. You have to build the 3D scenes and uh, sometimes stuff wouldn't even work 'cause you didn't shoot on green, so you'd have to rotoscope manually. Right. Yeah, yeah. Which is a pain and yeah. Yeah. I don't know how long it took him to do the, do this once he had the workflow built, but I imagine not long. Yeah. You know, probably a week or two to like just you, you know, there's a lot of shock. To shot stuff to tweak. But yeah, for Frame.io, like the biggest advantage here is not necessarily the background generation, but the fact that he's just shooting himself anywhere he pleases. He, yeah. It's like as long as the, it's like no screen stuff he touches. Yeah. And the angle is right. Yeah. That's it. That's it. That's it. That's all he needs. Yeah. He is not really recording camera tracking information and that the lighting, you know, kind of is like, yeah, but you can relight, you can, right. Yeah. That's more wor, I mean, now we're getting back into the fix it and post like, yeah. Issue. Yeah. Yeah. Well, it's, it's what we joked about a few episodes back. Like, just put the camera in the corner, put the camera anywhere. Right. Just film everything. And then you'll just reposition your shot and we'll just, we'll just redo it in post. Yeah. Cool. I thought it was cool tutorial. I think it's cool. Cool workflow. Good work maker for a filmmaker. Come on, come on the pause. All right. And then, uh, our last grouping, we've got a collection of stories of. Music and AI and AI in the music world. Yeah, there's a lot going on there. And uh, traditionally we don't really cover music all too much, but since there's so much going on, we thought we'd take a peek. I think that, I think also the music industry is good to look at because usually what happens there is like a step ahead of we'll trickle down into for sure. Into ours. Yeah, for sure. All right, so which, which one do you wanna tackle first? Because there are a couple of AI music bands. Making a big kind of stories. Yeah. So the one that fascinates me is, uh, there's a British creator called Oliver McCann with no musical background. He was a graphic designer. Graphic designer. Right. And then he discovered, I guess, Suno or one of the, yeah. Musical tools to make this, the music and his Suno Play had like 3 million plays and now he has a record deal. Yeah. Now he got signed with a record deal with, uh, Geffen like records. Right? Like a what would, what would actual musicians have to say about that? I'm sure they're not happy. Well, I mean, also, I mean, I, part of it was like, wait, this feels like a, this feels like an onion article 'cause as an ABC news article, and then his quote is. I have no musical talent at all. I can't sing, I can't play instruments, and I have no musical background at all. That's like all of us, except for like 1% of the population. But yeah, he, uh, started messing around with Suno and one of his tracks racked up 3 million streams. Damn. And now he is been signed as a AI music creator. Yeah. So this is where the world's going. No, I'm just joking. Look, I think this is a fluke. I think this is, this level of success with not having any musical background is not, it's like the equivalent of the Blair Witch Project to being like a total worldwide phenomena from like a$40,000 short film that somebody shot. Like these things don't happen all the time, and nor will they. This feels different ish in the sense of like, this feels like a weird fluke and this very specific moment in time. Yeah. That won't really be replicated because it's like, why do you need a person to do this? If you're going down the AI music route, what do you, what do you need a person for? Uh, behind this to make it right, but I mean, what like, that, that is an actual person and like, you know him as the brand. It's like you could just automate generate this. Oh no. I mean, the auten authenticity of that person and the fact that he doesn't know music adds to the lore. I will say there was a quote, I'm trying to find it 'cause it did say that the Suno songs, the lyrics are, are, you know, bad and that he, uh, uh, oh here, uh, quote, uh, AI lyrics tend to come out quite cliche and quite boring. Yeah. And that he often writes his own lyrics and creates up to a hundred different versions of a song before satisfaction. So look, that does. So yeah, he's doing some work. I mean, going through a hundred different versions of a song is creativity and that's input. And I mean, actually that, that is important though because a record label wouldn't sign him if he can't copyright protect the music. Right. And from that copyright, uh, guide, uh, at the beginning of the year, it was humans need to be involved. And part of the process for it to be, uh, protectable. It can't just be like it's generates spit out. And you're done. So if he's writing lyrics right, and then going back and forth with Stu, know the human involvement makes that, as we understand it now, it makes that copyright protectable. Yeah. This, this is interesting. You know, that is like, oh, we're signing him as like a creator. Whereas, okay, so the other stories, well, let's tell many the other stories because the other stories are there are AI music bands that are Yeah. Completely fake on Spotify. Completely. Right. We got the one on Spotify. Yeah. The Velvet Sundown. Yes. Is the band on Spotify that's fully synthetic? Exactly. And an indie rock band that appears to be entirely AI generated is making alarming amounts of money on Spotify. Uh, they've got fake, uh, first of all, band images. It's a whole fake, nobody makes money on Spotify. Snoop Dogg came out and said he never made any money on Spotify. Yeah. But yeah, they came out as like a quote, real band, and then it was revealed like, oh. Like the photos, the characters, like the whole thing is like a made up story and the music's generated, I think because it's gimmicky, it's successful, but if you weigh it against real artists, it's a novelty. Now that's, it's, it's capitalizing the novelty now. Yeah. But. Yeah. That, that, that's what I'm saying. It's like when, um, when Vine came out and all those Vine Tiktoks got big, like there was a first generation of 'em.'cause like it was a new type of Yeah. Platform, new type of content. Origin story of the, the Paul brothers. Yeah. So all those guys, right? Yeah. Uh, Lily Singh and, oh, was she originally? I think so. Like, so we're having that moment now with Suno and AI music. Let me put it to you in this perspective, would this type of success happen with a filmmaker who just texts to, generates a film and then a Warner Brothers buys it for 10 million? Not Warner Brothers, but it's like, I mean, we're sort of seeing this with, uh, you know, the AI shorts or the generated clips, like the weird AI stuff that keeps circulating of like the fruit that looks like a baby and the baby fruit is eating itself and the. There's just like all sorts of weird AI creations. Yeah. I would not, like, I, I'm saying they rack up views on social media. The, the kangaroo one, no one is signing them. The kangaroo. Right. Like that was a viral Yeah. You know, that one was more in the realm of you weren't quite sure if it was real or not. Right. The other ones are like, clearly they're fake because they're. No way. They could be real, but they're just like bizarro, like tap into that kind of as MR market. Yeah. Of like weird clip that people watch, but it's not it Warner, to your point, Warner Brothers is not signing them. Yeah. So only in music. Yeah. In the music world feels much more experimental. Like they're more risk taking than the media entertainment world. That we're in. I mean, yeah, maybe it's like, oh, we can ride cap some lightning in a bottle and ride this for a few months. That's it. Until my thought was just like, well. Spotify's working on their own AI music generator stuff where it's like, what? What do you need a person in the loop for making this when you could just have something just make music on demand. Especially if it is more of the background music kind of stuff. Like the lo-fi genre. Sure. Or the sleep genre. Sure. Or the relaxation genre where it's like you just make turn infinite loop, turn that out and have little, I mean, Google has some experimental tools where you could just adjust little knobs for like, what do you want it to sound like? And just. Keeps generating lo-fi beat music like forever. Yeah. Forever. I could, I could imagine like, uh, like an elevator company is what? Otis. Otis is an elevator company. Yeah. Like they would build in a GPU into every elevator. So you do local inference of a suno model. So you just have elevator music. Just generating on the elevator is not even connected to the cloud. I think that's too, I don't think, you don't think that's gonna happen. I think that's too expensive and too, too high tech for like such a specific meet, like an elevator company. I, when was the last time you've heard elevator music? All the time where I've never, hotels I have, like I, I've, you go to like Vegas hotels, they would have that besides like a Vegas hotel. What was the last, I can't think of any, like office building or, I mean residential. I've not heard. Elevator music, lobbies. Lobbies have a lot of music. Uh, yeah. Yeah. Like, uh, hospital lobbies, doctors lobbies. No, I mean, lobbies do. Yeah. But I mean, elevators, I mean, it's la We don't have any elevators, lobby spaces. Yeah. No. I have a friend who actually programs, they program the playlist for,'cause like a lot of these, like lar like restaurants or lobbies. Yeah. They need fresh music. They need playlists. Yeah. Like also lobbies are looking for not just like generic background music, but. Actual like, like the real songs song of the moment. Yeah, yeah. That programs it in. Yeah. But like, they'll build out, they have a whole system where it's like a playlist for the morning, playlist for the evening and stuff that transitions. And so it's like the whole Yeah. Day and environment and like locations are programmed and built out with playlists. Uh, okay, fine. Uh, maybe I went a little overboard there, but can I come back down to reality? All right, here's, here's my take on Spotify building their a AI, an infinite generation thing. Dude, nobody will listen to that crap. I don't know about that. Come on, man. If you're, well, I mean, if you're in the group where you're just like, I just need some background stuff, like, I mean, I'll load up Spotify's, uh, you know, the DJ lo-fi beats? No, I don't use the DJ. Hey Joey, what's up? You're back. No, I got tired of that. Um, I just use playlists, but I'll like load up Spotify generated playlist of like background beat music and uh, you know, like Lofi or something. And I'm not looking at who, unless it's like a really good song. I'm like, oh, I kinda like that. I'll save it. But like. 90% of the songs, like I don't, I'm not looking at the name, I don't know who it is. It's just there for like work music. But that's a curation algorithm, right? It is curated of other music. Yeah. But I don't think it's a big stretch for Spotify to just generate, generate stuff that sounds like that. That's playing like your next track is not this. Artist, but it's just generating, it's just generating the sounds. Yeah. But then you would listen to that maybe. I mean, some of the stuff on Suno is not that bad for like background, for like some videos and stuff, but I don't know. I just feel like life is too short to listen to synthetic stuff and look at synthetic films like we are here to consume other humans arts, and that's a big moment for us when we do that. Right. I think it depends on the, I think it depends on the the need. The use case or like what use case that thing is filling is, is fair enough going for right now? Yeah. At that moment that you're trying to do. That's true. Like, I mean the, the, the lo-fi girl, do they even say who the music is? I only know her as Lofi girl, but it's just always playing. It's probably an entire team of people, but I don't think they even like, I'm trying to think, do they, I don't think they like display, you know, like box tv, like the, the song title and, and creator, I think it, it's just music on loop forever. But like, so anyone, lofi girl, super popular except, did she like graduate or something? Did they like kill the channel? I think they said the FI girl, you know, graduated. Do you know what I'm talking about? Yeah. Okay. But anyway, it's just playing music forever. But you don't really know what the song is they're playing. Sure. So like in that same lo-fi girl genre? Yeah. You could just, Spotify could just generate background music and um, where Lo-Fi girl. Like the video or the track is being played in so many places you wouldn't otherwise expect. Right? Like in in an Uber or something. That kind of music. Yeah. Yeah. Like, uh, Waymo's, uh, yeah, they've got like just this chill spa music plane. Right. Uh, or you could pick other music, but it's like start default the spa. And it's just like as you're driving, it's a robot's driving you around town and you're freaking out that it's like they can not gonna like turn into a car. You're in an active police car chase that's playing chill music. Obviously sleep music's another huge genre too, and it's like you're literally falling asleep, so you're not gonna look at who the artist is. Yeah. And then, uh, when it comes to like kids and babies to put them to sleep, you need some type of lullaby music and Yeah. I'm sure there's a million used. That's what I said, like I think, I think there is room for both. Okay. Uh, you know, human and gen look, I mean, obviously Spotify was trying to figure this out because it's easier to build the algorithm to play the music. Yeah. Instead of paying people. Yeah. It's passive income for them. They're like, why not? We could just generate Right. Cheaper than cheaper than the royalties. Right. For better or worse. Okay. Last one. Uh, Higgsfield our favorite company. They also have, uh, their first AI idol. It is an Ai k-pop idol. And the same thing. They built a whole, like, you know, using, I assume Higgs Field's tools, uh, Kion Music generator. Yeah. Music videos. Fake character, the whole thing. This is like the Lil Miquela type model. Sure. For, for K-pop. A K-pop, yeah. Thing. But then they're also like, Hey, we wanna make more models, like, you know, be our next like AI model. Yeah. Wouldn't be funny if somebody steals her. I know. We should. We should. Yeah. I feel like anything that Higgsfield does is on the table to, uh, copy, replicate and run with it. So yeah. I mean, this is. An emerging genre of like, music's gonna be, you know, I think first to be impacted by it. It like, is it just gonna be a new genre in music? I mean, does it fit in safe spaces? I, after the Lil Miquela experiment with, you know, unreal engine based avatar, we've had a few mild hits. I don't see this as a thing, like there will be attempts, there will be many more kons. I think, but you might be too old for, we might be too old. We might be, you're right. Like you're right for, you know what the target demo, I take that back and to like that, where this would resonate or where they would not care. Like your kids might not care at all. Like, but, but the thing that with the younger generation, I'm finding is they're yearning for, and I'm super overgeneralizing here. Hmm. Um. In general, the reason why nineties and eighties aesthetics are back, they yearn for a time without connectivity with analog stuff. That's why they're going back to wired headphones. They're looking at Kodak, Kodak, gotta wait, wait to see the photo. So along that same sense, I think they will desire real connection with the real artist versus Just the stuff that meta did. Initially remembered like the whole AI generated family that's posting photos, that's, uh, yeah. And now the weird like, um, chatbots, like as much as we despise it, I think the younger generation will despise it more because they will be much more in inundated by it than we will. So I think, I think there's a correction period that's already happening. And with more AI slop, you'll have a harder correction. Yeah. And it's like, how do you do an in-person event for your AI K-pop store? Like how do you make. Things that you can, you know, go to connect to other people. Yeah. Because concerts make money, right? Concerts make a lot. Yeah. I mean, yeah. So is it just a giant TV screen if you have no talent and it's, it's a LED wall. Right. But I mean, uh, if you're, you know, your AI Kpop star, you do not have to pay them a bunch of money because they're not real. Uh, so like the margins are better. It with a, you know, if you're just are streaming and building out an online, uh, portfolio Yeah. Of an a IK pop star, but. You don't, you know, there's no, the, the connection you could have with an AI character is limited, like we talked about the little Taylor stuff. Yeah. The, the flip side of that is there is a small percentage of population who will absolutely connect with Kion. Like, uh, a synthetic avatar. And they've, you know, there are people that connect to characters that don't exist at all. Like a Luke Skywalker, right? Yeah, for sure. Yeah. Or, yeah. I mean, the whole world of anime and stuff and anime characters, I mean, but then a lot of times there's people go to. Conventions. Yeah. And connect with other people who also like the same thing. Yes. So there's still that there, there will be a community around communal element. Yeah. Uh, and stuff. And yeah. Last thing, this just goes back to like what I think we've been saying before, where it's like if you are a creator in the world where like AI automation is like, or AI. Generation is creeping in. Like I think that's where having a personal brand or like having like something about yourself that you are a person that people can connect to. Yeah. Is gonna be the differentiator for sure. Uh, besides just the actual work itself, because like, yeah, it's getting so easy for like the output to be generated or copied or spit out. But if it's like. Something that people can connect to like a person. I think that's gonna be the, yeah, I mean it, the buffer to like, you know, survive when like everything could just be generated in two seconds. In a world where we're gonna be drowning in a sea of noise, essentially finding something authentic and connecting to it is gonna be the thing that we. Yearn for hopeful. Hopefully it's as beautiful as that. If, you know, I, I, I hope we're not going into a world where, you know, it's just the, the commonplace thing to do is to just have an AI friend or go to an AI concert, or look at an AI movie. Maybe that's not the world. That I wanna live in. I, I know, I, I, I think it'll be fine. I mean, look, look at, you know, the record numbers of people turned out for the Minecraft movie and the, the Minecraft experience and, uh, the K-pop, uh, demon hunters. Yeah. You know, huge. Like, yeah, because Oh, so found they went to the music. They went to go party and hang out. Netflix's biggest movie to date. Yeah. They just passed. The most viewed, yeah. Yeah. Box office for the weekend, last weekend? Yeah. I think they, uh, they uh, average it by number of hours viewed and or number of viewers. I think it's at 237 million at the moment. Yeah. Much higher than anything else. Yeah. Yeah. Uh, I hope that it's a musical and has songs 'cause the stuff with songs. Yeah. Has that repeatability factor? Oh yeah. They, I think they're already working on a sequel. Yeah, I saw that. Yeah. But yeah, look at the popularity of that. People went out to the theater to like go hang out with their friends and like, you know, go sing along at a musical. Right. So I think, I think human element of like going to stuff is hopefully, will always be there. Great. Great note to end on. Good point to end on. All right. That's not what you think. Link for everything we talked about at Denoisedpodcast.com. If you wanna go back to our previous episodes on YouTube, uh, we have tons of them, I think over 60 of them. So give some of our older episodes a watch and uh, comment on it. Thanks everyone. We'll catch you in the next episode.