Adobe MAX 2025: 15 AI Updates That Actually Matter for Filmmakers

0:00Welcome back to the noise in this episode we're going to cover all the new product updates from adobe max patty hello joey let's get it please report to room seven joey are you uh am i player four five seven

0:23i still happy halloween i forgot to wear my inflatable alien costume that i did buy that I'll wear tomorrow. Awesome. I still have not seen season three yet. I still need to catch up on that one. Season three sets us nicely for the next season. So definitely give that a go. I thought it was the last season. I thought it was it. Is there more? Come on. It's just getting started. There's more Squid Games. Yes. All right. So WMAX is happening right now. We're wrapped up by the time this has aired. There have been a ton of new product updates. a lot of stuff we've talked about before, a lot of stuff that I feel like we're seeing integrated into actual products.

1:05We can just run through everything that we've been seeing online. Yeah, sounds good. Let's give it a go. So in the kind of first press update that they posted, grab bag of AI updates. So Adobe Firefly, the model, because this always gets confusing now because there's Adobe Firefly sort of the product, which is like their AI platform. And then there's Adobe Firefly, their generative, trained on commercially safe data model. The model is now upgraded to version 5. And from some stuff I've been seeing online, a lot of the improvements in that, way better with people and portraits. But this is the model that's used under the hood in the features on Photoshop, like Harmonize, which we talked about a few episodes ago, which was you can kind of add a couple layers into Adobe, and then it just blends the lighting and blends the subject into the background more.

1:55That's using Firefly under the hood. I had seen people say that it didn't work great with people, but that Adobe Firefly 5 improves a lot of that, and so it should work better when harmonizing with people. Yeah, my guess is they just retrained it with more data, right? One of the biggest lack of quality came from the fact that it was licensed data. And as you know, there's not nearly as much licensed data as there is publicly available data. So I'm hoping and guessing that Firefly 5 is just maybe in the billions of parameters, perhaps from the millions. Yeah, that is a tongue twister, Firefly 5.

2:34Firefly 5.5 on Halloween. Also some other integrations. This was integrated, but I think it was a beta version of Photoshop. But basically Photoshop had, in addition to Firefly, obviously, but they added Nano Banana and Flux, Context as models you could call up. I think that's now in the official release. That was like a beta release version before. But they also added a generative upscale with an integration with Topaz. And this is across, I believe, built into Photoshop and also as an option in Firefly and Firefly boards. So that's also cool to see they have other upscaling options, you know, with their partnership with Topaz as well.

3:12Yeah, congrats to the Topaz team. I saw some of the pictures of them at max. I was like, oh, yeah, there you go. That's an ecosystem building right there. Yeah, they've really been having, I mean, they've really been having a good moment with like stuff they've been doing for years is really the timing has worked out with the rise of AI materials. And also, I mean, even if you're not doing AI stuff, upscaling is useful to have just across the board, especially if you can get good, sharp details. upscaled yeah the traditional upscaler that's built into photoshop for example is still using um math right like you know interpolating between pixels and uh i think it's called like bicubic interpolation stuff like that uh whereas the topaz one will actually use an ai engine yeah i uh yeah i remember asking banji uh when he was talking about astra their generative video upscaler um Like what, how, what, like, so like what was upscaling before?

4:08And it was sort of just like, it's complicated. Don't worry about how it worked before. Like it's math. It's a lot of math. Like, dude, you're not going to just, just use it. Yeah. But yeah, the generative model is it's like they train their own model, but that model is on like taking blurry images and making them sharper. And so it's a generative. I mean, just like thinking super long-term and I think Topaz knows this as well. when the image models are fundamentally not going to be a small pixel size anymore, like when they're actually natively 2K or 4K or 8K within latent space, you're not going to need upscalers.

4:49I mean, maybe not as much, but yeah. Yeah, I mean, I'm curious to see when that happens and if that's the case. But there's still, you know, if you crop it, if you're like maybe you do make a generation, but you want to keep something but crop in or zoom into it. you know, upscalers can come into play that way. Yeah, or maybe upscalers will evolve to upscale bit depth and dynamic range and some of the other things other than pixels. Right. I mean, we saw that workflow from ETC where they used, they didn't use any of Topaz's generative models. Those were all their math models. But basically it was like a workflow to take the 8-bit AI images and turn them into something where you can get more latitude out of them using a combination of like Nix and Gaia, I think, and some of their other models to denoise it and then add some more dynamic range into the images.

5:44I mean, it could still also be cheaper and faster, even if we do have models that are spitting out AK, to generate something faster at a lower res and get what you like and then upscale it. And it might be cheaper and faster to do that still than to generate multiple iterations from the start at 4K or 8K. Yeah. In the end of the day, a lot of filmmakers will tell you, like, it doesn't matter how you got to it. It's just that you got to it. That's the other updates that they renounced. This one's a public beta AI object mask in Premiere. Automatically identifies and isolates people and objects in video frames so they can be edited and tracked without the need for manual retoscoping.

6:22Wow. This is cool. That stuff like that is fun. Yeah. Wait, did you say we don't have the rotoscope? I'm in. But yeah, not even that, but also just like click and just detect things. Yeah. Yeah. Awesome. Jumping back to Firefly Foundry, which we talked about in the last one of the last episodes. Yeah, we covered Firefly Foundry and the Invoke acquisition on two episodes ago, I believe. Yeah. And so Foundry was a way to, you could, I think I got some of this wrong because it was initially posted as like a enterprise level custom way to build your use firefly as as a foundation and train your own model on top of firefly using your own assets your own ip your own material and have your own i haven't used it yet but my guess is it's a self-service where you can build your own laura so for example you know i'm wearing this beautiful squid game outfit if i I wanted myself to wear this outfit all the time.

7:21I would take a bunch of pictures of me wearing this, and then I would build a Laura, and then I can call preset that every time I generate something of my likeness. Yeah, and it seems like that is an option. So it's called Firefly Custom Models. I'm still thinking this might be two services or might be Firefly Foundry, which is like enterprise will train on your products or something a bit more bespoke. And this one is Firefly Custom Models. It's in a waitlist version now, but this is similar to what you said. Self-serve, upload your own images in the style that you're looking for, and you'll have your own version of Firefly that you can call up in Firefly or Firefly boards.

8:01The Foundry stuff feels more enterprise-oriented. Like, that's more for companies to use, where this one feels more consumer-oriented for you and I to use. Yeah, this feels like mid-journey mood board feature, where you can give it a bunch of images, and then you get an SRF code, and then you can generate your output, but give it your custom SREF code that's trained on your collection of a couple images and the style that you want. So it's like a much easier way to direct and navigate the output that you want. But what's cool about this is it's still using Firefly, so you still have that commercially safe foundation if that is what you need to generate with.

8:39It's super important for a lot of people, yeah. And would you say this is more one-to-one with Flux context now? I would say this could even go further than that because flex context, there's no, you can't like, I mean, I guess if you trained a Laura on that, but that's a bit complicated. Like this is like a very easy way to, and not even have to know what a Laura is to just be like, this is the style that I like. And I want to make more images in the style and you give it this group of images and then you don't have to worry about prompting or styles as much. You just kind of say what you want, and then it has that under-the-hood Laura directing and navigating the output to give you something in the style that you want.

9:20In that sense, I would say it's maybe even better than Nano Banana because it just is able to retain consistency over a long period of time. You don't have to, like, throw reference images every time that you generate something. Yeah. I think these are different use cases. I think it's like FluxContext and Nano Banana. Those are good where you're like, I need to, like, direct and make very specific changes to an image. and I want to specifically modify a shot. Whereas the Firefly, the custom Firefly model, or whatever it's called. Yeah. Whereas with the Firefly custom, it is literally called Firefly custom models.

9:53Yes. It is like, I want to, I have a style, like even the examples here, you know, it's like a very surreal flowery style with, you know, whales floating in the sky. And you're like, I have a very visual style that I just want all my outputs to look like this style, that's what you would use this for. And I think this is an expansion of, because inside Firefly already, there is a group of preset styles, like Sketch Illustration and 3D Isometric. They're preset styles, and this seems like an expansion of that where it's like you could just make your own style to call up to generate images in a certain style.

10:33So different use cases. Like, this is not a... It's like a style transfer on steroids. Yeah, exactly. So it just saves you this hassle of having to try to do very specific prompts every time to get a certain style output. Sure. Different use case from like Nada Banana or Flux Context. Okay. Then it's more like runway presets. Yeah. Yeah. Runway presets, mid-journey mood boards, training your own Laura. Pick your favorite analogy. Whatever. There's a lot for you to choose from. Just go and use it. The next on my list was from Theoretically Media. There's a lot of just people in the audience posting screenshots and stuff from the keynote, so getting a lot of info from there.

11:15A shot of called Project Graphs, which is a node-based AI editor, which could already be Invoke that we talked about from the acquisition last episode, rolled into Adobe in some fashion. I mean, that would be really quick if that integration already happened. And again, this just could be like that. I mean, you think they're independently developing their own graph products, and then they happen to buy it. I mean, yeah, probably makes sense. Or it could be that this is a very containerized one-off demo just for Adobe Max, and then Vogue team just built it just so they could put it up on screen, right?

11:58Like, this is not a problem. Yeah, or just to also put graphs on people's radar or node-based editing on people's radar. You know, we've still talked about how, like, is Adobe and Adobe users, are they the node-based audience? Are they going to adapt to that? So maybe it's, like, easing, you know, if someone's, like, very much in the Adobe ecosystem, node-based editing is probably a pretty new concept. Like, we wouldn't have really been exposed to that. So maybe it's sort of just trying to, like, ease people into that. Like, hey, nodes and graphs, it's a concept. Hey, Adobe user, you've been using Adobe for 20 years.

12:32Now look at these. Yeah, here's another option for you. Yeah, no, Adobe's never historically had any node-based anything, right? Like nothing I can think of. No, we talked about this, yeah, before. I can't, everything is layer-based, even going back to, I mean, obviously After Effects, but even going back to like Flash. Flash was layer-based. Yeah, all of their web building stuff, like Dreamweaver and things. Dreamweaver. Yeah. Yeah, nothing that I can think of that was node-based there. Yeah. So, yeah, it's a new concept. And I will say, a lot of these demos, I forgot what they called them, but, because I was on the Adobe livestream last week, and so that was a cool experience, and the team was great, and they mentioned, because we were just talking about, like, what was coming up at Macs or other features, it was like a category of things, like, they share a lot of stuff of just, like, test concepts.

13:23I forgot, there was a word for it, but it was sort of like, you know, like, concept cars where it's like you see the cool car but like it's never actually going to go into production it's made out of clay ideas yeah adobe has like you know groupings of these sort of like lab projects where sometimes they might actually make their way into production and actually into a product and sometimes they're more concepts or things that they're experimenting with that what you see demoed is not actually ever going to ship it's more of like a work in progress or what they're thinking of like if you remember two years ago at neb they were demoing like like generative AI built inside Premiere, you know, with like the extend the scene, which that has shipped.

14:04But then they had another concept demo that was like, call up any model you want inside Premiere to generate a new video clip. That has not shipped. I have not, then that was like a year and a half ago that they demoed that. So, you know, there are features and stuff that they'll demo that might never actually go into production. Yeah, it's so difficult for somebody like Adobe who makes software at scale, right? Like they're making this stuff for millions of users across many, many different industry verticals. We tend to think of only film and TV, but, you know, Adobe is fairly in play in automotive or, you know, fashion design and everything else.

14:42Right. Like for them, it's so challenging to invest in a product, roll that product out and then have it flop. Right. Like, so how do you get that initial feedback from the market? And I'm sure they do siloed user testing and things like that, but that's still not like, you know, rolling it out to 10 ,000, 100 ,000 people and getting that initial feedback. So, yeah, I feel their pain. You know, of course, you want their tool sets to get better over time, but then what direction do you take? Which products do you invest in? And so on. Yeah. And they have such a massive user base where it's like obviously impossible to like please everyone.

15:23But when you have a specific tool that's so widely adopted, you know, like Photoshop or Illustrator, that, yeah, any one change could anger a lot of user base. Well, yeah, that's, I think they have pleased everyone. Like, they're one of the few companies that have a universally accepted tool set across all creative industries, right? Like, you think about Photoshop, Lightroom, even Premiere, that's used across the board, many different places. Like, who else is a competitor to that? I would say AutoCAD on Autodesk's side, like that's used across universally as well. Like there's very few players like that.

15:58You know, we're talking about the creative projects and stuff, but there's also so many other business aspects of Adobe that like are so B2B enterprise-y. Like they have like a whole, I think they have a whole like separate like CRM marketing. Yeah, it's called Digital Experience Cloud. I think internally it's called DX. So you and I, we're the consumer side of things. So we use creative cloud. Then they have the digital experience cloud. And then they also have another segment of the company called digital media experience DME. So the other two is where I think the bulk of the money is made on the enterprise side.

16:37yeah yeah so i'm saying that basically they're just like outside of creative mac well whatever else max is talking about but like creative cloud and or creative suite there's like other massive business verticals in adobe that are like yeah and this is a company with like i'm gonna say five billion in revenue every year like it's a massive company so again it just makes it that much harder for them to experiment with new things when the thing that they make works perfectly fine yeah and you don't want to yeah yeah you don't want to rock keep innovating but also like not annoy your base customers that rely on your products every day yeah all right other new stuff this one's cool this one's a good example of just like practical ai under the hood stuff that makes things easier.

17:23Layered image editing. So basically, if you import an image into, I don't know if this was tested in Photoshop or just a Firefly feature, it automatically segments the layers and then identifies and labels the layers. Oh, wow. AI-powered layer labeling. AI layer generator. Yeah, so here, this image of a, and generative fill, too. It's like they added an image of a soup bowl. What do they call it? Layered image editing. Layered image editing. Their names are very literal nowadays. I know. Yeah. Let's just be clear, not clever. I'm just thinking for virtual production, how useful that would be for putting up plates and getting parallax out.

18:08Oh, just moving stuff around. Oh, or that too. Oh, yeah. Like automatic separation of the mountain from the tree from the foreground. This is awesome. But look, it separated the chopsticks and then as you're moving them, it generates the missing parts in the background wow that's useful because one as you know i i'm like a you know intermediate photoshop user i'm not really good but one of the things that i always have a hard time i'm rotoscoping in photoshop right like i'm using the pen tool to like separate out the layers so this would save me tons of time yeah uh so yeah this is a good post from uh alexandru costine who oh is the vp of adobe firefly there you go all right thanks so i guess Yeah, it's a Firefly product.

18:49I see they tagged Adobe Express, too, which is sort of like their... Yeah, that's their... Canva. Canva web-based... Yeah. Yeah, I use Adobe Express. It's, I mean, gosh, I hate to say it. It's not nearly as elegant as Canva, but it definitely feels very familiar if you're familiar with Photoshop. Okay, yeah, that makes sense. I also saw there's like a Photoshop cloud version. I've not used these tools in a while because I've just been in the video world. But, yeah, I was exposed on the live stream because they use Adobe Express a lot for building out graphics and stuff. And it seemed powerful enough for a web tool to build out social media graphics.

19:26Yeah, I'm building out, like, birthday invites and, like, random graphics things all the time. So Adobe Express is perfect for that stuff. All right. Other thing that ties into something that reminded me of a thing we talked about with Facebook a while ago. Project Moonlight, Adobe's AI social media admin. and it basically looks at your existing social media assets and then starts generating new social media assets. Wow. Wow. That's cool. There's a lot of competition in this space for sure. Yeah. I mean, I think it's tough to compete against Facebook when they're also the platform. Yeah, Meta.

20:05When they're also the platform as well. Meta is not like the friendliest company to compete with. they'll all right just steal your stuff right like they'll just build this into instagram yeah this is cool i mean it's interesting just like and this expands the idea of just like okay we want to personalize make a bunch of assets social media assets at scale that still fits our brand and our style yeah how can we have the ai learn and like keep generating new stuff from that so we can just kind of keep posting new fresh content that still feels on brand dude we should just stop doing this podcast and just have moonlight take over yeah we could just with talking uh we'll just get uh synesthesia why can't i speak today whatever the other the ai avatar company is we'll just i think we have enough footage already we just yeah we have enough training we have enough training data although my quick game suit is messing up the data today oh yeah that's true you're throwing it off today all right let me go back to adobe press release uh new ai creation tools generate soundtrack this is in public beta an ai music generator powered by Adobe's commercially saved Firefly audio model.

21:12That's new. Generate speech, text-to-speech tool for creating voiceovers, also powered by Firefly. But they also did announce a partnership with 11 Labs. Oh, that's huge. As well, as a partner model. Dude, 11 Labs is absolutely killing it in the audio game. Yeah, I mean, they're pretty much the standard for, like, decent generative voiceovers. Yeah. Let's see. This other beta, Firefly Video Editor. A fast, intuitive, web-based, multi-track timeline editor for generating, organized, trimming, and sequencing clips and tools to add voiceover soundtracks and titles. Oh, interesting. I would have to take that for a spin.

21:48I mean, somebody like you would probably take a look at it and be like, ah, it's not good enough. I mean, reading this description, this sounds like something for social media clips. You know, sounds like an auto-edit for creating 30 - or 60-second reels or something. How does this compete with something like Descript? I mean, I'd have to see how much control you have after the fact. But yeah, Descript has similar features like auto-edit and some auto-editing features. But then it still also has this pretty powerful text-based, web-based video editor. So you can tweak and change stuff and have control over that.

22:27I'd be curious with this Firefly video editor, like how much control you have after you get its initial first cut at whatever it was trying to do i mean if the cuts are generated in premiere then it's all like non-destructive right like you can then join the clips again and take over essentially yeah i mean i would assume this is non-destructive as well but just how much control do you have yeah like i'm assuming it's something like oh let me give it a couple like you know my raw footage i shot on my iphone and edit it into something interesting that we can post on our social media channel um i mean it sounds like it also kind of competes with premiere the iphone app which they just launched um because it's like that was also sort of trying to be built for like mobile creators yeah we talked about feed io and some of the other competitors in that space yeah um you know and this could be one of the things where they're just testing the waters maybe they'll integrate this into premiere sure uh it's labeled as a private beta right now so like who knows where you know where that goes i'm curious to see how the generate speech quality compares to like vo3 which is not the best you know oh well i think i don't even this is just text to speech this isn't even like they're putting speech oh lip sync firefly video outputs this is a the generate speech is a competitor or just a complementary tool to uh 11 labs got it okay i haven't seen them say anything yet that they're generating a speech with the with the firefly video what's the chinese model that is now competing with vo on both audio and video one yeah one one two point five yeah right yeah okay yeah i still think vo3 is the best at maybe maybe sore two is getting up there but um yeah vo3 is probably still the best with generating video with matching and lip sync and all that yeah the synthetic human health just doesn't interest me enough so i don't look into it enough because I fundamentally believe that a human performance is the center of any AI film and that AI is just there to do essentially VFX, like put them into a background, put foreground elements on, put smoke, fire, whatever.

24:36But the core performance is not a synthetic human that's lip synced and VO and all that. But that's just me. And I know that the capabilities definitely out there and people are going to use it to make films. I'm glad that there are more tools rising that you can take a human performance and use that to drive synthetic characters. For sure. 1.2.2 Animate is the perfect example of that. This one's cool. Project Light Touch. This is a clip posted by Howard Pinsky, Principal Design Evangelist at Adobe. This sort of reminds me a bit of Beeble. Basically the relighting. import an image yeah i don't know if it just only works on images or if it does video to the demos with image yeah and then he's adding spherical light sources and relighting the image based on the position of the spherical light i welcome this in a heartbeat because relighting other than bbl i can't think of any other platform that's really doing it and this is something that a lot of vfx artists are asking for actively because relighting in the traditional sense with image based lighting and HDRs is just so cumbersome.

25:48Yeah. Yeah. And this demo, I think it's just still photos. Obviously people does full on video. But yeah, and this could, this is a demo project who knows what they would actually ship or integrate this in maybe. Yeah. Again, like, I think the point of Max is to kind of show you glimpses of what all of the R &D is going on inside this massive company. Even as like 50 % of this stuff actually makes it out and turns it into, a product and features that we could use, that's a massive win for all the users out there. I got a bunch of clips from Jared Liu, who is an AI educator, and he got a bunch of good shots from the front row at Max.

26:28Why can't I be an AI educator? You can't be. Just put it in your bio. Just change your bio. You're right. This one, project frame forward, lets you alter the start frame, but keep the same motion of the original video. I don't get it. can you explain that this demo clip i don't think is really showing it and it also looks like he's working in photoshop let me see so it's people doing yoga on a paddleboard oh okay so it seems like generating let's see what this new output it'll keep the video as the same video but then you can in paint within it yeah so in this example they erased one of the paddleboarders but kept everything else the same.

27:10So it's sort of like a, you know, like a Luma's modified video or runway olive-ish. I mean, this took a lot more work to get this modified, but if it didn't mess up the other two people at all, that's a pretty good test. I can't see in this video if the other people are like super consistent and if it's also the same original resolution as the video, because the issues I've had with like Luma Modify Video and Runway Olive is, you know, if you do something like that where you're like, I want to erase a person, the output video, it tends to warp people. Yeah, it slightly damages the original frame, right?

27:51Yeah, and then your output quality is going to be less than HD. Yeah, so it's like, do I really want to go this route because you're degrading the quality and you're going into a different avenue? so that yeah and i think that's pretty useful i'm sure they do yeah obviously that's an issue um and makes it less usable for actual professional applications uh if this solves that that'd be cool that's cool uh this also sort of reminds me of um or just the way that he described it reminds me of a feature that is in moon valley where you can have your starting frame and a video to drive the animation.

28:31A simpler example of this would be like, what if your starting frame was like a basketball and then you have a very simple blender animation of a sphere just moving around. And then that drives the animation of your ball. Very simple example, but you could do some, you could take that with what you want of like starting frame and driving video. Yeah, I think when we covered Greg Teagarden's work, he was able to do the same thing with comfy ui as well like give it a motion reference video and then have the starting frame follow that motion reference yeah yeah yeah i think we're seeing that we're seeing that more all right another update project surface swap a tool in photoshop that can automatically detect surfaces to make fast adjustments let's see this demo So they selected a car.

29:24I mean, this is not the best demo because you could change, replace color in Photoshop traditionally. Yeah, you could just select a hue range, right? They could have picked a mountain in the back or something more complex. Right. But if you're not a color theory expert, and this is as simple as clicking the car, and it automatically detects all the colors of the car and blends it. basically does this make something that would have taken a lot of tweaking into something that's now like a click yeah there's like a recurring theme here along i mean the colors look good and it blends it looks like it's blending really well like it doesn't yeah and it's keeping the um like the highlights and the shadows well too within that color range yeah i mean it's obviously everything they're going to show in a demo in front of an audience is gonna look good yeah You'd hope so.