Nano Banana Pro, Meta's New SAM 3D, and World Labs!

0:28All right, welcome back to the noise. waiting to make a left turn all right so big news this week nano banana two but it's actually called nano banana pro well technically called gemini 3h3 or something but everyone knows it by nano banana and they're calling it pro interesting naming convention which we'll get into all right so biggest updates well what for one you can now do 4k resolution outputs which the only other model that i know of on that par is c dream c dream 4k which has been great just having that extra resolution. I've been testing those two side by side. Yeah. You notice anything off the bat?

1:01I mean, look, viewers, let's just face it. We got the best image generation model to date. And we're going to talk about it today. I'm super excited. Yeah. I mean, the biggest thing, just crazy diagrams and text is on next level. And I've seen cases of people giving it research papers. I mean, make a diagram explaining this. And it's like, make a big diagram. So it has a reasoning model built into it. Yeah. Because also the other update from Google this week, which Gemini 3 came out. And so you have an under the hood, you know, these multimodal models. So it's not just a denoising image model.

1:33It is understanding the world and turning it into an image. Yeah. And it's got probably a lot of LLM engineering built into it and it's plugging directly into the logic from Gemini. Yeah. I mean, I'll say right here in their blog post with Gemini's three advanced reasoning, not a banana pro doesn't just create beautiful images. It also helps you create more helpful content. So yeah, it's using it under the hood. I saw a crazy example where somebody took a gundam you know the action figure toys yeah and said uh give me this as um as like an assembly like a toy thing on those little sprockets oh and it did it it just like disassembled it in each parts and components yeah accurate diagrams like make a flow chart here this one like make a recipe a chai recipe it's like images text text is rendered super crisp been clear and accurate it's not gibberish it's so versatile yeah however i was just testing what i know and love which is film and tv applications and even then it outshined nano banana one significantly oh yeah like what app what things did you see yeah when you were testing it we're we're gonna do you want to get into the yeah let me just the other thing like so let's see uh text i mean so yeah text is a big update also character consistency has been has improved and we'll probably talk about that more yeah the uh film applications elements so you could do elements before where you give it a couple images and kind of it makes yeah whatever you want now you could give it up to 14 images and it says now bena pro you can blend more elements use up to 14 images and maintaining the consistency and resemblance of up to five people so five characters it could handle and understand and be consistent which is also also the fact that they have that number and are clear like it's five people it it should perform well with consistency with five yeah and i would say that's an understatement some of the examples that i've seen online you're looking at 10 plus people so imagine you and your family of 10 you put everybody's photos in there and say you know put addy in the middle this person on the left this person it'll just just make it that's crazy and then obviously like reframing and spatial understanding and that's where i tested it too of just being giving an image and being like reframe here reframe there which it was not a banana original was good at before it's now even better and absolutely more consistent more accurate i've got some tests too though that i'll run i'd love to see your test yeah i think spatial understanding of 3d environments without being a 3d model or i don't actually know how it's architected under the hood yeah but it has almost good enough spatial understanding to be world accurate like as far as units go maybe yeah and i mean as we talk about these are all leading to world models physical ai understand the world and can accurately reproduce and recreate and behave in it i'm gonna say your line today okay yeah what all things lead to robots all words lead to robots so you did some tests and you did whiplash test so what's this test we're looking yeah so i just picked one of my favorite movies of all time whiplash and you know jk simmons is most well-known role i would say so yeah i took the iconic frame of jk simmons at the opera house and then i wanted to do a background swap it sounds really easy to do and most models can do it but really it's in the fine details and the subtleness of how well it does so my prompt specifically asked for relighting in a cloudy environment so on the top right i have the Seedream example.

5:02On the bottom left, I have Nano Banana 1. And then bottom right, I have Nano Banana Pro. Now, JK Simmons, like me, is a ball guy. And as ball guys, we have usually a hotspot where the light hits us, right? We're like a giant chrome ball, if you will, HDR map. So when you light somebody from a direct light source to an indirect light source, that hotspot should go away. Right. And that doesn't happen in Nano Banana 1 or Seedream. However, in the Pro, as you can see here, he's being lit by like a very diffused light source, aka clouds. Yeah, yeah. I mean, look at this. The Seedream, it got the environment, but the lighting looks pretty similar to the lighting from the original where he's backlit, very warm, harsh lighting.

5:47The glare is on his head. Yeah. Nano Banana 1, a little softer, but also his face was warped a little bit. So it's also not as consistent. and then yeah Natumana Pro just nailed it just nailed it like his pose his face is exactly the same but the glare on his head is gone the lighting feels like it's an overcast cloud lighting yeah and the fact that it's a natively 2k model and cannot output up to 4k I think that just intrinsically helps with small details and it sees more it has more resolution to our eyes we just pick up on that stuff so much more and to us we feel that it's more authentic yeah yeah this is a good side by side test okay all right next one i got another image all right so i was trying to invoke a little bit of the cantina from space uh from star wars okay yeah so like you know uh like a cowboy saloon space style and i was like okay what would it do with all the reflective materials in there like the symbol how about the bandmates in the back and i just want to see how this will perform a style change or style transfer it makes this entire scene a space sci-fi opera without changing the two actors in the foreground alter the instruments to be more sci-fi but keep him in the same place yeah do not alter actor poses or camera perspective only relight them cool this is moody indoor and dark yep so i'm gonna just skip to nano banana one did a decent enough job but then the symbol has this like weird electricity across it yeah and the stuff in the foreground doesn't really match the background like the lighting kind of matches it's not really relit whereas with nano banana pro it took a whole different approach to lighting it completely relit the whole thing everybody has this purple tinge to them the reflections on the symbols are so accurate even put a little bit of an outfit change on jk simmons which i didn't ask for but it gave him a little little robo flair yeah yeah and and like a little padded like a judge dread type-esque outfit yeah he is the judge dread of uh band band directors yeah and and again it's just so much more realistic like it doesn't have the neon accents that seed dream has up top yeah you know a synth pad too and because you said change the instruments make it more futuristic right yeah yeah so again very impressive tests and then let's go to the last one okay so uh this is uh miles teller's love interest i forget her name look i just want to now try the visual prompting here so i use free pick and then you could see my examples i wanted her to be in a starbucks in a barista outfit with espresso machine in front of her but i didn't want to change the pose i didn't want to change the um sort of the moody pose that she has i wanted to keep the original intent and guess what nanobanana one completely changed it uh so she is uh in a different camera perspective different framing and slightly different face she's not behind a bar yeah her face changed a bit she looks a bit more serious and uh nanobanana pro kept that ambiguous mood to her and you know it's it's very sort of what how what am i feeling it's a little complex and yet everything around her changed so realistically like you have the syrup bottles in the back the espresso machine i'm trying to look at the menu too because like also the text on the menu and i'm not even sure if i can't read it because just the image you saved is lower resolution or because it does not look like the normal gibberish that you would get from a from like AI outputs.

9:12It looks like actual font. That's just, I can't read it because it's like out of focus. Yeah, and I didn't give it Starbucks reference images. I just - No, you just said Starbucks. Right. Yeah. The other thing here that you'll notice - Logo is accurate. That you'll notice is she's also relit correctly. I think the relighting stuff is really just impresses me more than anything else just because that's such a hard problem to solve in the world of VFX. Yeah, yeah. I mean, it did keep the same window motivation because we see the same light on the espresso machine, light on her face that feels consistent yeah i mean like six months ago we were looking at anabanana one outputs and we were thinking boy that is so realistic and now you have the pro output next to the original and you're like that's so much better yeah i will say you know and i've seen some other examples of this there is like a little like two percent difference on her because here it's like her lip is a little agape and you know it looks a little bit more longing and then in the output it closed her mouth yeah made a little bit more serious right Right.

10:09So it is reinterpreting her and regenerating her. But the fact that it could get to like at 98 % close. Right. Let's talk about little details here. Right. Yeah, this is a good test. So one of the issues I've had with Nano Banana or just any model that has spatial understanding is if I want to get a reverse angle. So like I have a shot of something and I'm trying to storyboard or just something else and be like, hey, give me the reverse angle of this. And I'm asking for the camera. I'm using film lingo, but I want the camera to do a 180. and show me the other side of the scene. Okay. And I have a lot of trouble.

10:45I've had a lot of trouble with that off and on. Sometimes it gets it, sometimes it doesn't. And then this one was the original image. This one was one where the test still did not work well. I've asked for, I wanted a reverse angle like over the dad's shoulder. So we're looking at the children's book. Sure. And in this case, it kind of just gave me the same, it moved the dad. Same angle. The same angle. So this one didn't work as well. and maybe if I tweak the prompt, it would work better. Or if you switch to visual prompting, perhaps, like place camera here. Perhaps, too. And also, Freepik introduced a new camera, a 3D camera feature, where it sort of turns the image into a block and you could reposition your cube.

11:22I saw that. And I'm guessing it uses that as an input. Yeah, I'm guessing that's an extension of NanoBananas API. I don't know if it's API, but maybe it's one of those things they realize, if you map an image to a cube and rotate the cube and you give that as an input it understands it that's maybe they figured that out yeah sure my my theory okay we can get Joaquin back on Joaquin's like you're so wrong once again so this one I said show me a reverse angle over the shoulder looking at the line of patrons the movie theater and it got pretty close but I mean it didn't get me over the shoulder in the sense of I was thinking like dirty over the shoulder so that it would be partly in frame but it flipped it her shirt that same pattern the hair is still in a ponytail it still has the same little purple markings on her shoulder and it's like looks like the opposite image of the movie theater i'm just impressed by the sheer number of people it generated because as you get as you exceed like three or four people ai tends to just mess up the rest of them yeah they're all relatively sharp i did and i did say looking at the line of patrons in the movie theater and there's a big crowd that's definitely a line so then i did i modified the the prompt a little bit and i said you know over the shoulder but the woman's shoulder should be right frame and we're looking over her left shoulder and it's like kind of partly in frame and then it it nailed it nailed it yeah yeah that's nice yeah so this has been the most successful model for posters like who asked for that i mean just whatever that's crazy isn't it's like big vacuum of its uh world understanding i'm curious what it's like looks kind of like what movies are those kind of looks like the hulk i think we're just looking into the matrix at this point like it's a simulation movies that play in the lane space in the theater.

13:04You're never going to get stuck there. You're going to just keep watching them in loops. I'm sure AMC would love to see a movie theater this crowded. Oh, ouch. After they did the Nicole Kidman thing. I don't know if that'll happen again. Yeah, they should all be protesting that they want the full 30 second Nicole Kidman spot back. Yeah, so that was my quick test with it. And you need to have a pretty specific prompt of what you want. But that's been the most successful test out of any of the models to be like, give me reverse angles of something in space. And what we talked about is because it naturally has a really good spatial understanding of spaces.

13:35What's the Google world model that we covered a few months back? Oh, Genie. Genie. Yeah, I'm wondering if there's some Genie guts in there. In the bottle? In the bottle, yeah. I don't know. Yeah, good question. Yeah. I mean, I imagine that, you know, these are all eventually heading towards just one. Just being one thing. Super world model, create anything that understands everything. I mean, I think architecture-wise, it'll still be a separate block, like the whole diffusion model thing is a separate block from the LLM. But how it's API'd and interacts with the rest of the world, it'll just be one giant black box.

14:12Gemini 3 or Gemini 5, by the time it gets there, will have a thing of Genie in there, a thing of Nana Banana in there, and thing of whatever in there. And this will also just be in other products. Like, I mean, we know about Nanobanana and we talk about it and stuff, but like just regular Google users, it's these are built into like now notebook LMS upgraded. If anyone used that Gemini 3 is powering it and Nanobanana images, they wouldn't know under the hood. They don't care. You know, yeah. If they're using Google Slides and they want to make an image, let's be like, wow, these images look way better or something.

14:41You know, Nanobanana is powering it, but it's built into the suite. So you wouldn't know. This is that paper diagram where this person copied the text from a paper and then told it to make a diagram and it made a whiteboard. It even looks like a real whiteboard. That's crazy. Yeah. These are some other. This is a roundup from Google. One shot at an entire detailed menu without a single spelling mistake for a nano banana fictional restaurant. Just little illustrations. That kind of looks like the island's menu. I think it could be. A little decorative and fun. uh scientific diagrams this one was fun a flow chart for how to toast bread make it as wacky and over the top and complicated as possible i love it it just looks yeah yeah like who would have time to do this in real life yeah it's like you could take your jokes and memes to the next level amazing you know uh joey i was just thinking about like that the ai race yeah and all the trillion dollar companies and open ai is just kind of falling behind at the moment so our two came out like a month ago and you're like sleeping on that step it up yeah i'd be curious i mean because um yeah for a while chat apt image was before nano banana they were like really good with the image generator the image generator with world understanding being able to make very specific modifications and reference insertions and stuff like that yeah yeah it's still good it just has that ai very synthetic look and you know the stuff out of nano banana is just looking photorealistic I heard that when you upload like a photo, like if I wanted to put Joey in a chat GPT generation and I take a photo of you uploaded in there, it'll intentionally not generate you 100 % just to avoid deep fake stuff.

16:20Yeah. So it's like just naturally dialing it down. Yeah. And I don't know what the safety features or limitations are with Nana Banana Pro because I have seen people say or I mean, they've made images, you know, of Trump. Of celebrities. Of celebrities. Yeah. that look for realistic obviously it is not really them yeah i haven't yeah i have i don't know what the limit is of censorship of what you can um make within anana banana pro i mean like a company as as safe as google like they've obviously navigated through a lot of legal hurdles over ai with the last few years i wouldn't think that would make a bad move like it's a very calculated thing that yeah they they're not like a startup that's like yeah we need to like build up uh they're not a field right yeah we need to be edgy to like build make a name for ourself right yeah the other thing too and i haven't really seen anyone post about this is the pricing so there is a bit of a different pricing as of right now so regular nato banana you make an image it is about four cents an image 0.039 yeah nato banana pro 4k 30 cents an image yes so about 10 10x i saw my credits just go yeah well also free pick and a lot of the other ai aggregators they're all kind of running unlimited generations with nano banana pro for the week okay so you could use that there free picks unlimited model has um added a lot of asterisks asterisks asterisks yes yes we we talked about this couple episodes back even more so now i think and i haven't dug into it but i think nano banana they cut down on to um yeah they've they've uh they're they're trimming the yeah i'm sure some people were taking advantage of that especially if you have something like manis like an agentic browser or something yeah oh yeah yeah keep churning away so i'm sure there was abuse of that but for now now i've been at unlimited for a week except also you can only generate one at a time yes because i in free pick i like to crank it up to like four or eight just give me all the options but with that one they're just like no just one do one and wait uh so yeah this is cool the um you know the one thing that you know i feel like it's on par probably better at some things than C Dream.

18:29But the one thing that C Dream does that I still haven't seen anyone else do is that batch generation in the same... In the same noise seed. Yeah. I'm wondering if they'll eventually do something like that too. Or if you're like, I need to create a couple of images out of the same realm and get even more consistency. Or they might say, what's the point? Because our reference images are super accurate. Just keep the reference image consistent and the outputs will be pretty consistent. when i was looking at your movie theory example the reverse shots i was thinking the same thing i was like yes it did reframe but the people are completely different so if it was the same seed and the same prompt potentially when you go from over the shoulder to the straight on that it's the same number of people well in their defense i didn't it was a reprompt from the original image i didn't oh i didn't take this image and modify it and say reframe oh gotcha It was your source image was the same.

19:26You re-rolled it. Yeah. I just changed the prompt, same reference image and ran it twice. Sure. So, in their defense, it was not, I did not try to reframe this image. Alrighty. Yeah. Okay. Next up, new paper model. No, it's a full model because you could use it now for meta. Sam 3D. This one I've been seeing going around a lot. I like the name. Very easy to remember. I know it stands for something. Don't ruin it for me. I know. Okay, anyways, Sam, give it an image. It will extract objects or people. There's two different models for that. And turn it into a full 3D model that apparently looks very, very, very good.

20:06Insane. Insane. Yeah. I mean, you know, just like a year or two ago, we were lucky to get image to a single 3D object done. Like if you have like this roller coaster thing, each of the cars would be like a single image. And then you would put it all together in a 3D environment yourself. and now it's giving you all of that in one shot so then you export it into blender or unreal or whatnot and you're off to off to the races yeah like i'm just looking at i'm looking at this picnic thing this food image and just like every i every individual piece of food turned into an object that's 3d yeah this is um this is in the i think the one of the last frontiers of generative AI is going to be 3D.

20:50Everything that we do in 3D today, right? Not just the modeling and the texturing, but also rigging, optimizing meshes. It looks like the people are also rigged too. Yeah. Oh, shoot. Yeah, it is placing a basic skeletal rig in there. So I take that back. I was going to say those things are coming. They're here. They're here. I mean, this is really impressive and cool from Meta, but it also makes sense for them since I don't think they have, you know, they're still pushing the metaverse 100 and this is like oh hey how can we make things more 3d faster yeah give it an existing 2d images and we turn it into 3d objects one of the biggest hurdles to the metaverse 3d world yeah yeah is is just getting the world populated yeah and if you're gonna give the users the ability to make their own worlds you can't expect those users to be 3d modelers in blender like they're not gonna yeah you're gonna wait forever for that stuff to build out exactly or you're gonna have to do something lo-fi like roblox or something well even with roblox roblox has roblox studio which is actually i tried using it it's pretty cumbersome to put a game together like really it can't be an average joe and do it you got to be pretty like on a blender level-ish level five out of ten in blender okay you know you got to understand meshes and rigs and stuff okay so like you got to want to do it yeah yeah right and fortnight uh fortnight islands which is the metaverse side of fortnight where you can build your own islands that's like a level 10 out of 10.

22:16Like you really need to be a game level designer. And so what a lot of production companies do. Like a special version of Unreal to do that? Or is it? Yes, they do have UEFN, Unreal Engine Fortnite, which is basically like, gosh, they're going to hate me for saying this. It's a stripped down version of the engine because a lot of the core engine features are not supported in Fortnite because the whole entire thing has to get like flattened down to a few gigabytes for it to publish. So you can't bring in the full photorealism of Unreal in there, like 8K textures. I mean, you can, but then you're going to go through that limit real quick.

22:52This is the complete antithesis of Roblox and Fortnite. It's like if I were to build a world in Horizon, Metaverse OS, then I'm just going to mess around with some generated photos, some generated people, put it into SAM 3D. It's going to come out rigged, And then I just dial in the game logic with a llama or whatever other engine they have. Yeah. Or I mean, yeah, llama's a bit slow, but yeah, imagine like, you know, taking the, you know, this blended with something like Google's Jenny and Gemini. Right. cause that was the other thing with Gemini three. It was a big improvement in coding, vibe coding, all that stuff, building out full interfaces.

23:35Uh, but in a gaming world, there was like, Oh, if you had the Sam three and it builds out all the objects in your world, and then something like Gemini, hey i wanted to you know i wanted this game to exist and then it builds all the interaction with your 3d objects or something crazy vibe 3d gaming vibe gamer yeah and then last one marble from world labs just launched so world labs has sort of been in beta papers for a bit and we've we've talked about i've had videos on the channel we've talked about them before but basically you could give it an image or a text prompt and it makes a very small navigable 3d world but enough where you can move around a bit and all of your objects are there if you move around too far it kind of falls apart but now they're actually released the full product it's called marble and a lot of i mean a all roads lead to robots and gaming obviously you know understanding the world from that case but you know for our sense uh their launch video was like a virtual production showcase so like they generated the worlds on a wall had actors interacting with them this is what we've talked about of like yes you know you could you know don't have to have a huge fad budget, you can generate your world and get enough where you could have reverse angles, some parallax, some camera movement and generate your scene.

24:44What a good use case to go after to showcase. It's a sexy one with the volume and everything. Yeah. And you can see it on the wall. I can actually have people there. You can have interactive lighting. You could export the scenes you generate as Gaussian splats or as rigged models, you know, bring it in Unreal. So a lot of use cases there, a lot of applications. And so it's cool that it's, you know, finally out in as a public product. How is the resolution or sort of the poly count, if you will? I mean, SAM 3D, from what I just saw, looked pretty good. And I don't know if this is maybe as good as that.

25:16I mean, it's tough to tell, too, because some of these demos are more in the game world. Yeah, they're stylized. They look like... And there is also an editor, too. You can edit the things that generate. So that's cool. You can give it an input image. You can do text to 3D. I mean, if you're building, let's say, like Grand Theft Auto for the metaverse, right? And you just need miles and miles of city, city blocks, restaurants, arcades, whatnot. What better way to do it than just let an AI agent go at it? Yeah. I mean, I feel like that's be like a kind of like the Jenny thing. You know, this feels like an extended use case of the House of David breakdown that we had.

25:54Oh, yeah. You know, where they were talking about they were it was a battle scene. It was desert. And they were just sort of generating 2D plates on the fly. Right. On a wall and shooting with that. Right. This is like the next step after that. Yeah. you know something where you can actually move around a little bit interactive a little bit interactive they have some camera have some parallax have a more consistent environment but you're drastically cutting your vat time down and that was that's the whole promise of unreal is you get the game interactivity with the photorealism needed for cinema yeah but it just takes a lot of time to build an unreal agency and this is you know it depends on the project depends what you're doing it depends on the project but if this makes you know having more locations accessible for smaller budget productions yeah great option exactly like uh you want to pull that unreal lever when it's absolutely necessary to spend a hundred thousand dollars on that scene yeah but for if it's a shot of five ten seconds long just you can get away with this or vertical short yes yeah all right good place to wrap it up awesome links for everything to talk about at denoisedpodcast.com all right y 'all so uh thank you again for our support on spotify apple podcast and youtube we have some new commenters i hope you guys don't mind if we shout you out we just love all of the comments and some of them are feedback which is great too we need it uh we have ms sharp 16 peter king day 77579 and misha buloi thank you for your comments hope to see you on the next video yeah you need to stop talking over me i know i know i'm so sorry yeah so it's not addy's fault the riverside delay sometimes happens but now we're here in person so we have we shouldn't have it as much crosstalk there it is a little bit my fault for some reason my internet sucks that is yeah it is the internet service provider issue i live in a part of la where they don't have fiber optic it's just too old school gotta get starlink or just move houses just so we don't have latency yeah all right thanks everyone we'll catch you in the next episode