Qwen Edit - The Free AI Image Editor That's Better Than Flux Kontext? Plus More AI in Film Updates

0:28All right. Welcome back to Denoised. screens you can kind of roll around and they're like well what if you could just you know talk to that and create with that and use that as like a super smart creative board yeah it's kind of the gist of where they're going but the funny thing is when we shot it the opening is with like a demo of uh kj saying like hey create an image of an office and when we shot it we didn't get a shot of like the final office image appearing on the board and it was really bothering me that we like didn't have this you know final hero shot so i'm like maybe we could generate it so i took a screenshot from the film and then i did the little vo3 trick of like marking it up of like what i want with like some text instructions yeah and then i gave it to vo3 yeah and it generated it no way and it's up there and then i didn't i i sent a a cut to tim uh the ceo and i'm like hey you know check out the video and then he watched it's like yeah that's awesome and then i'm like anything seem weird to you yeah he's like no i'm like ass one of the shots is from vo3 he's like no away yeah go check the video out and uh if you see the shot you know leave a comment uh let me know what the shot what shot you think it is look out for it now you're gonna be looking for it but if you watched it before cv that's awesome that's such a real world use of ai saving the day yeah an insert shot you didn't get you're also a super practical man you're a stickler you need the pay i do that shot you needed that shot like it's really bothering me we like have the setup we have like no payoff reveal of what the thing was created that's awesome yeah so uh put that in the toolkit.

1:53Things that can work. The whole sketching and then like writing notes on that frame. Yeah. It works. Yeah. It's so interesting to see some of these kind of tricks that people discovered and that they're able to apply. So like VO3, but I think some other models, people realize, oh, you can kind of do the same trick with that too. And then also another sort of recent trick is specifically with C-Dream, which we've talked about a bit. By dance. By dance model that I really like people realize that in the prompt if you tell it like change shot to something else and then change shot to in the one generation you can make like three different shots but because you're in the same generation you get that that world consistency right and you get a couple shots out of the same wow uh starting frame environment yeah same lighting yeah same which actually reminded me of that thing we talked about from that hackathon a few months ago with uh dylan our man dylan and that trick where he was just generating in the same generation like doing a bunch of shots and it was like he was he was onto this a while he was brute forcing it and then now you can just do it yeah now it's a lot more cleaner looks great do we need dylan anymore sorry dylan dylan gpt just joking yeah so and then i think some other models have realized like oh you can kind of do the same thing in this model too so it's interesting like one model figures something out or someone figures out a hack and then yeah other people adapt it and test it out with other models and you realize oh you can kind of do this too oh that's so cool yeah All right.

3:13So first update, probably the biggest one I've seen this week is Quen. The image model we talked about last week came out from Alibaba, open source model. They now released Quen Image Edit, which is basically another version of the model, open source, but you can speak or give prompts and modify an existing image. Sort of like a flux context, but this is open source. You can run it on your computer, make changes, and kind of have your own portable. Go into a comfy workflow. There is a comfy native workflow already built that you can like load up yourself and download the models or you can run this off file or any or replicate or any of the other APIs.

3:49But yeah, often we see something that is like flux context level for free, for free, for open source. Again, as we talked about last week, he's like, what's the what's the play here? Why is this free and open source? But yeah, yeah, pretty impressive. As we're wondering what the Chinese AI company's secret game plan is, and I have some tinfoil hat theories on this. Every week, it's just like, boom, boom, boom, boom. And this is a 20 billion parameter model. Yeah. You know how expensive that is to train? It must cost millions of dollars, thousands of GPUs. I mean, they do. It's Alibaba. They've got the GPUs, right?

4:28You still need the electricity, the engineers. Dedicate the time to run it. To train it and watch it. And each training takes months. So they started on this probably end of last year or something. Yeah. It's crazy. Yeah. Or your tin hat. Oh, you guys want to hear it? You want to hear it? Okay. All right. I'm going to lose some viewers on this one, but I think it's worth it. Now I want to hear what this is. I think the Chinese are bombarding us with advanced AI models. So we lose the ability to make our own high-end models and we become more and more reliant on China. This is a big geopolitical play.

5:11Is this like flood the market with like cheap manufacturing equivalent? Yeah, and it worked, right? Now we're trying to bring manufacturing back to the US, but we can't. Like, oh, I haven't built a plastic toy for three decades. Toys R Us used to buy it from Minnesota back in the 70s. But now the Chinese have made it for the last 30 years. So now what do I do? I think they are just making sure that all of our consumer and enterprise AI needs are met from their models. And eventually our AI muscle will atrophy. That's my theory. Come on. What do you think? I don't know. Yeah. I mean, my only thought was from before was like, get people used to and hooked into this and their workflows and then switch to a charging model or switch to.

5:59I don't think they care about the pennies that you make. I mean, honestly, they don't. Like, they'll never make the money back on Quay. No, I don't know. I don't know. Yeah. So the other... I mean, aside from just a geopolitical play of just like, we would rather the world be dependent on Chinese models. You got to also remember, China is always playing the long game. They're thinking 100 years out, right? And because AI is such a disruptive and transformative technology, they know America now has the upper hand. The best models are still here. and like open AI is here, Google is here and so on.

6:30But then if they do this for long enough time, then the open AI business model kind of breaks. The Google stuff, they transition out of AI, go back into search engines or whatever they do. And they just use the Quinn models, the Alibaba, the Tencent models. And maybe that's how they sort of make sure that they still have the upper hand. Or this is the only way to sort of compete in a noisy space is to like make it free. Sort of what Meta was trying to do with Llama, where they like were kind of late to the AI game and then they built Llama. And they're like, we'll just make it open source. Like, hey, use it, adapt it, because they're too late to compete with like ChatGPT.

7:03Also, meta doesn't need the money from Llama. Exactly. That's, I think it's a combination of maybe both of those things. Yeah, just the only way to play is make it free. Yeah. Yeah, I'm messing around with it a bit. Definitely on par with Flux Context. And I don't know how it compares. I mean, I've seen some test demos comparing it to Not a Banana, which we talked about last week. But it was on LM Arena, I think. I mean, when we talked about it, but I went to LM Arena. and I can't find it there. So I don't know if they took it off. And I don't know how people are playing with it. Nano Banana? Yeah, it was on LM Arena as a model you could test.

7:33Yeah, I was showing you the website, nanobanana.org. No, you told me there was a nanobanana.org. I'm still, I think, I'm not convinced of that as a real nano banana. Oh, I paid for it, man. I think you were - Where did my money go? I think you're getting fluxed. I'm going to guess it's the same thing as a nanobanana.ai, just another opportunist like spun up website. Wait, what? I showed everybody nanobanana outputs and I got fluxed? I mean, what, did you find it? than using flux i didn't do an a b comparison i mean just in your general experience i don't know like did you i i thought it was highly impressive uh yeah i've used flux context in the past um i thought it was equivalent maybe not okay maybe not better well this website is saying that it is better than flux context look i mean if this is really a model from google that is not on their api or anything then how would this website be using nano banana that looks like every other because it's in limited review right like you website from what i from my research it sounds like Nano Banana is out and in limited preview.

8:28So limited preview to NanoBanana.org. They're the ones that got access to it. Damn it. I think I got spoofed. I'm curious. Also, if you run NanoBanana.org, you can, you know, reach out and let us know if we're totally wrong. But yeah, I'm suspicious of everything now. You got more tinfoil hats than I do. Maybe I do. I mean, look, if it's truly from Google, and I was hoping we maybe we would get some of the, you know, reveals at the Google event on Wednesday around the Pixel camera, but nothing in the image generation model there. I don't think I didn't rewatch it or anything. Yeah. So you think the Nano Banana stuff will tie into the Pixel phone?

9:07No. I mean, I was thinking maybe they would announce something with it, but they didn't. And there was just a phone update. Yeah. And then the whole like Jimmy Fallon thing. I've only seen blurbs about that. I didn't really like dig into it. Cringe. Yeah. Was it cringe? It was pretty cringe. Sorry, Jimmy Fallon. the only commentary i saw was people saying like i i miss apple's live presentations but after seeing this i understand why they are not doing it live anymore yeah yeah yeah i mean like i don't watch it but kudos for them for doing something live yeah like the whole like late night show is such an outdated format anyway like all those guys i mean stephen colbert retired or got pushed out yeah it was canceled uh conan has his own thing now jimmy fallon any day now and then seth myers successfully transitioned to podcasting right yeah right and he had like a late night show on youtube for a while or or not youtube tbs it was on tbs after that whole jay leno fiasco yes yeah yeah so like end of a end of an era yeah i mean i don't really watch late night i mean the only thing is um you know if you see clips or something that pops up on youtube and and uh yeah i think jimmy fallon the last one to kind of hold on to it is saturday night live like they're still somehow that's different too i mean it's once a week you know late night it's a skit show yeah it's a skit show it's commentary it's once a week and they always bring in up big celebrities plays well for clips but yeah i mean late night late night is four nights a week and it's like oh i know i'm not keeping up like if i'm watching something that consistent it's something else yeah like alien earth super good dude i so okay so i tried to sign up i signed back up to hulu the other day and it was tied to a my disney account yes and then i i was like i have a disney plus account i put my email in couldn't find me and now i'm like in this weird uh zone where my login has disappeared from disney plus and also hulu oh okay i mean i think you could probably access it through disney plus because they're sunsetting hulu eventually and they're just rolling it all into disney plus eventually i should get a disney plus account again and then i think i might be wrong i think everything that's currently on hulu i think you can currently access it on disney plus if not now that is the future they're okay rolling they're killing hulu and just going to move it all into Disney plus yeah which the screenshots of seeing stuff like saw six like thumbnail with the Disney plus like icon on top is really it's really funny right next to Mickey Mouse Clubhouse so yeah you know but I think it makes sense for long-term play yeah I can't let you have all the fun man I mean you were uh ahead of me on the studio and then I caught up studio so good oh your recommendations are spot on I'm gonna watch I try to do I try to just pick the hits i mean you know if you're into uh you know downtown abbey kind of stuff the gilded age is also great oh that's gonna be a tough sell you know i'm just saying yeah gilded age sure awesome show have you seen the gentleman the show on netflix yeah better than the movie yeah i love the show the show was great and then i watched actually never saw the movie i watched the show and then i watched the movie and i'm like oh man i'm glad i watched the show first because the show is way better right okay yeah all right the show is awesome good soundtrack too all right let's moving on AI guys we're turning this into a TV show TV commentary talk show all right back to AI stuff uh runway had a bunch of updates I'd say like little bits of updates that collectively are like oh they dropped a lot of stuff this week but nothing like wow yeah crazy first one they added voices to act two so act two their motion capture system you can record a performance of an actual person translate it to uh an AI or just an image character before it would just be whatever audio you had in your video would just transfer over now you could restyle that audio as well only to like a limited selection of voices they have.

12:38I feel like this would probably be better if they connected it to 11 Labs or something, which might be in the future because we'll talk about it in a second. But yeah, so you could change voice, limited selection of voices. Yeah, I think the key here is the timing of the voice and not necessarily the voice itself because you can always go from audio to audio model like 11 Labs, put the runway voice in there and then prompt for the, like, I want a deep voice. Right. And then get the right voice out, bring it into Adobe's SoundSuite. and then add reverb to it or what have you. Right. I think the main thing would be doing that route is making sure that the timing stays the same because you want the lip sync to...

13:15Yeah, so I think if it comes from runway, just like, you know, VO timing is good, but VO quality is not good. Do you agree? VO timing is good. So when VO3 generates audio... Their audio, yeah. Like, lip sync is perfect. Yeah. But then it sounds like AI. Yeah, yeah. There's always, like, a little weirdness to it. Yeah, it's like... Yeah, I mean, they win in that category. are the only one who will also generate audio in the video right now yeah that's the advantage of going through the runway workflow is that timing and performance will match that's the hard you just have a limited selection of voices you can pick or if you if you're really an advanced sort of post-production person then you can put it into um an adr session and then just put down your voice or somebody like a professional's voice over that yeah and i think that's what people were doing before because you were stuck with whatever audio was recorded on the video yeah sound is so especially voice voice is so far behind from video generation or you do animation style the reverse you record the audio or generate the audio first and then yeah lip sync your performance to it and then run that through love that runway yeah yeah old school animation techniques coming back okay and then the other thing with runway a couple of updates is the big one was they are now opening the platform to other models so not just runway models but the big one is they uh integrated vo3 as a model in the platform select third-party models will now be available directly within chat mode allowing you to choose between a more robust set of pipelines to better accommodate your specific needs so that's interesting because yeah sometimes i find like runway and runway olive works great for some stuff not so great for other stuff and i think in this kind of ethos of like everyone's trying to be the one-stop platform so you have to do less yeah model hopping they're trying to build a wall garden yeah yeah i mean runways had a pretty good interface too yeah you buy a thousand dollars worth of credits in runway how do you spend it all give you more ways to spend it yeah yeah without having to jump around to other to other systems right but that's why i'm also saying like maybe that opens up the door for like an 11 labs integration because i definitely have the best uh voice models and so you know maybe the oh i'm sure christopher was looking at 11 how could he not but yeah goes to runway that seems to be the most buzzy audio model at the moment like i've heard so many other people talk about it too not just us on the show no i think 11 apps has been the like standard for any sound related stuff where it's like been the best voice quality audio quality of the real-time audio that now they have music they have sound effects yeah they've definitely done pretty well established themselves as like the audio ai yes leader and the last runway update runway game worlds which i believe we talked about a while ago they kind of demoed it as a private beta which was it's not a immersive 3d world like we've been covering with uh like world labs or uh genie 3 it is more of a game world like text-based world generator of like more of like creating the dynamics of the world of a game and it's more about generating the world itself and the characters and the dynamics but more in a text or image-based environment so more of like a game mechanic engine so anyways that is now out of beta uh people can start messing around and playing yeah i don't think i can get my mind's around what is what is the offering so it's not a world generation it is a it is the world generation in the sense of like another is the need for novel mechanics and interfaces yes this is this is good uh and this is strictly in what makes unreal engine so great yeah go on yeah it's not it's not so much the generation of the world yes unreal is really good at that but it's the blueprints yeah it's a world map being it's logic trees and decision trees and then um having multiple levels having the scoreboard and player stats reward systems like more of like this the scripting phase or the planning phase the actual game design itself is is 90 that and 10 visuals yeah yeah so that's what they're building with game worlds i mean i only imagine that this will eventually translate into videos and something you can move around and explore and as soon as you have real-time image generation real-time video generation like this stuff is going to be ready by then yeah and then these are the mechanics and these are like you know the the custom built unique experience that you as a gamer might want to play and experience i think uh with the real-time generation it won't be that you know the cloud is doing all of the generation and then we're pixel streaming down to our phone it i think generation will get efficient enough and the phones will get high performance enough to where it's doing local inference on the phone.

17:41Yeah. And so, like, let's say, you know, we're going to get the iPhone 17 this year. By the time the iPhone 20 comes out in a couple of years, this stuff is ready, and now you're playing AI games on your phone. But the average, you know, Gen Alpha won't even know it. It's like, oh, this is just a cool game. Yeah. And we're just like, whoa, that thing is running AI. Everyone has one of their little NVIDIA AI computers. Yes, the DGX, Sparks. Yeah. Yeah. I mean, imagine that miniaturized down to a mobile phone. Like you can take the guts of that and I'm sure put it into Android phones. Yeah, eventually.

18:16Yeah, for sure. Generating on the fly. And the latent space. You know, it's funny when we talk about hardware, I was listening to a proper news article, which I rarely do. I'm very opinion based. But there is just such a massive shortage in silicon right now across the world. like it's not just nvidia can't make enough gpus but you know tsmc that taiwanese silicon manufacturer like in order for them to make a chip it's like 31 000 steps it's in it's so crazy too there's like only one company in one spot of the world that can make all of these chips yeah and and um like because the the chips have come down to like such a small size i think we're down to like six nanometers for a single piece of wire that's like a thousand times thinner than a human hair or something like that that's crazy that at that level of complexity there is only one company that could do it and so like you know we went through the the automotive chips during the pandemic right so many cars were just sitting because they didn't have chips on them and now we don't have enough chips for the gpus and there's just chips missing for physical ai so like in order to build a robot you don't have enough like uh embedded chips to go in them because of that i think we're going to shift to more efficient models by just the sheer necessity i mean it makes also it's like it makes sense too i mean and from even from a energy point or just resource point it's like with the deep sea argument too it's like well you need to train the big models first yeah to then be able to like shrink them down into smaller more efficient models yeah you still need to distill them down yeah you need to have the big training first for that for the second part yeah and the training i don't think it ever get more efficient right now i mean we're 20 billion parameters today we're going to get to 200 billion parameters very soon and there's no way around more gpus like you got to use more gpus but the inference i think there's a ton of improvement in just a local inference and we're going to see some really exciting things in the near future yeah especially when it's just like regular kind of everyday stuff you need to do yeah you don't need to like solve complex right physics problems on the go yeah yeah right right you just want something you could talk to and you know can like handle things for you yeah like i want to make fun of my brother all the time and it takes so long to generate a video of him being fat because he's very fat that that is what you want to use your silicon for yes silicon shortage i can't make fun of my brother chip here we go all right next update uh this is a new integration on file but um i hadn't heard this model before and i thought it was kind of interesting it's called mirror low sfx and so it is a audio sound effect generator no prompt you just give it the video it looks at the video and then generates what it thinks the background sound should be interesting kind of more targeted for synthetic ai generated content that doesn't have sound i got tested with a couple shots of like some people on boats and it made like the wind blowing and the how good was it waves i mean it was like definitely better than nothing okay and good like i would add a couple more elements to it but as a baseline foundation good starting point okay and like foley yeah it just lets some background you know it was like as if you know it was like your on camera audio was recording like some background sound oh okay so yeah i thought it was a you know i mean no one sound effects and sound design for ai and ai generated stuff is something not really talked about or covered right definitely i mean look sound is super important and anytime i watch a lot of the you know ai content and then you turn the sound on and then the sound design kind of falls apart or adds to the uncanny valleyness of the whole thing if you have really good sound you can trick the brain or cover up a lot of weird looking visual things yeah yeah i mean one of the reasons why star wars the original one still hold up is because of the like also many of these icons or right because those sounds are just etched into our brains the sound design on that show was so good yeah yeah and you believe it yeah your brain believes it right that this is real yeah even though they're just holding completely sticks and you know little little uh models and a black box that they shot.

22:18Right. Yeah. So yeah, I thought this was... One of the sort of technical marvels that's running under the hood for this, as well as the runway voices on Act 2, is a model called VLM. We typically talk about LLMs a lot, large language models, but we don't talk enough about visual language models. So it works in reverse of image diffusion. So you can give it an image and it'll detect features and objects within it to a very precise degree and then feed that into an LLM. It's like this, you know. So what would you use that for? So in this case, if you upload a video of a boat in an ocean, that's the systems that's detecting the ocean, the boat, the waves, the velocity of it and all of that.

23:03And then telling an LLM, this is what sound should generate. And then that's going to an audio model. Then it's generating the sound. Okay. And this also ties into a model being a multimodal model where you can give it text input or image input. I think by now we can assume that all the models are multimodal. They're so advanced. ChatGPT is multimodal. Yeah, I think all. Maybe not video. I think Gemini can handle video input. I don't know if ChatGPT or... I should try that. I don't think it can. I think Gemini can, but usually if you give it a video, it's not looking at the whole video. It's like one frame a second.

23:40To me, image generation, video generation are like cousins. I mean, so much similarity between the two. I just kind of bucketed as one thing. Audio is another bucket and so on. So you can think of VLMs as the eyes of the AI system. So I would imagine if you have a physical AI, like a robot, the camera inputs go into a VLM and that's how it's detecting a person and this and a camera and so on. That's how it'll see the world. It's already doing it. I sent you a video last night. Did you take a look? No, I don't watch that one. That was, what was it, robots? You're scared. You're scared. what does he send me so it's always people messing with the robots so this robot's trying to pick up these objects from the box and this guy has like a golf club out of all things and he's just pulling the box away from the robot and the robot's pulling it back they're gonna remember this i know it's so mean all right other this one you know kind of more just grab back a quick updates uh 11 labs music now has an api so you can plug it into your systems comfy systems whatever oh amazing and generate music through however you want so handy to do that and then uh last one that i saw this one is kind of a quick update but um i actually think this would be pretty handy google gemini will now read your google docs out loud so this is actually kind of useful because like i uh you know especially i'm driving a lot over here and so welcome to la yeah so uh you know sometimes if there's like a long article or like some document or i'm like i don't have time to read this, but I'm going to be in the car.

25:08Like 11 Labs actually has this separate dedicated app where you can give it documents and it will read it in like an 11 Labs voice. And so it sounds really good. So I'll use that for documents sometimes, but just also having an ability if I have a Google Doc and I'm just like, hey, I'm busier doing something else. Like just read it out loud to have a pretty good Gemini voice realistically read a long document to me while I'm doing something else is actually a pretty useful feature. Yeah, that is. Do you know if this works on the mobile Google Doc app? I don't know. So if it does, then that's ready for the car.

25:39Yeah. I mean, I imagine, I mean, everything's so mobile focused. I imagine if not in this release. Probably a couple of releases from now. Yeah, in the future. Yeah. For me, I'm like a convenience for, or a sucker for convenience, right? So like, if I have to like generate the audio file, then load the audio file into something and then play. Yeah, no. And then also, then I just have all these random files in like Apple Music. And I'm like, no, no, don't. We're not doing that. No, I don't want that. Because then I'll be like, hit shuffle. And then it's like music, music. And there's like some random audio clip.

26:08And it's like, what is this? Yeah. No, no, no, no separate files. Yeah. Okay. I mean, you're right. This is very useful. And I think we covered that. Was it Google that generated a podcast? Yeah, that's Notebook LM. Notebook LM does that. Yeah, they do that. Yeah, that one is a, yeah, a little like that one will actually take the data and then synthesize it into a two person fake podcast explaining the concept. and they launched a video version semi-recently but the video version that i tested was more like it built out a fake powerpoint presentation it wasn't that interesting because it was like i don't really want to watch power like i don't want to watch powerpoint presentation of this document i gave you like just to be honest when you're sitting on the 405 you might there's nothing else to do yeah for those of us that don't have autopilot on our car and like actually have to pay attention it's hard to watch a video oh oh yeah sorry so uh tesla fsd 14 is dropping it's supposed to be a big game changer and autonomous driving what will that do elon said quote-unquote it's sentient okay it's supposed to be a big game changer i'll let you know how it goes all right yeah i told you this earlier a way my mom almost sideswiped me today you live in waymo city though yeah waymos are everywhere but yeah we don't have the tesla taxi thing yet robo taxi yeah that's not in la yet that's in san francisco the bay area yeah oh we don't get anything out here i would try that purely from the price point because the waymos are are not cheap they're they're on par with like regular ride sharing well i mean the they're expensive the the lidar systems the software research the ai models like somebody's got to pay for all that i think it's more demandish i think they're just oh their search pricing yeah and i've launched the app and it's been like prices are higher than usual because demand is high and it's like i heard you book it through the uber app or there's a separate can there's a way there's a dedicated waymo app okay yeah all right i'll have to try one of these days yeah all right enough filler in this episode too yeah sorry folks it was a light week all right we're just kind of riffing that's the end of the week man we gave you nano banana first before anybody we did cover nano banana first you know and uh maybe maybe quen image edit will be on your radar too yeah until we can until nano banana the real nano will the real nano banana please stand up all right that was good that was real good we'll end it there uh thanks for everything we talked about at denoisepodcast.com.

28:25Shout out to OLE92 for leaving us a wonderful comment on Spotify. Thank you for your support. Thanks, everyone. We'll catch you in the next episode.