- il y a 21 heures
Human Creativity to the exponential power of AI
Catégorie
🤖
TechnologieTranscription
07:31L'AI agent est plus capable, plus fun, so smart.
07:35Essentiellement, vous êtes remplacé par l'AI agent.
07:38Mais, let's get practical.
07:40Et, let's break down
07:42comment on think about l'AI agents at Adobe.
07:44Dans notre vue, les agents ne vont pas remplacer vous.
07:48Ils vont vous reflecter.
07:50Think of them comme un créatif-helper
07:52qui s'exécute efficacement ce que vous envieziez.
07:56Not bad, right?
07:57Who couldn't use an extra junior creative
08:00on their team for a project?
08:02Imagine the agent as a hyper-competent assistant.
08:06Someone who dialogues with you in a conversation,
08:09takes creative direction,
08:11works with you to refine the project based on your feedback.
08:15You could also grant this agent access to all the tools and data
08:20that you use in your workflow
08:22to make sure it's all seamlessly connected.
08:25We're still early in our development of our agentic approach,
08:29so let me show you one agent that's already available
08:33and two that are in development.
08:38If you're in video, you'll know this Chrome.
08:41It's Premiere Pro.
08:42Media intelligence is already available in Adobe Premiere Pro.
08:47This is our film editing application.
08:49If you're a video pro, you have had countless opportunities
08:54of searching through hours and hours of footage
08:59looking for that perfect shot for post, right, in post-editing.
09:04Now, with media intelligence, you can scan all the footage,
09:08tag it, metadata tag it,
09:10and intelligently define what's in every shot.
09:13That way, when you need footage,
09:16you can search via the agent and find it instantly.
09:19That's hours saved in post.
09:21In this case, they were looking for a mouth-watering tortilla
09:25with tomatoes and lettuce.
09:28Okay.
09:29We're also working on, in Photoshop, an intelligent agent.
09:34And the concept here is the agent will analyze your photo
09:38and proactively suggest ways to improve it.
09:41In this one, a lovely hero shot in Paris.
09:45There were those pesky tourists in the background, right?
09:48Go figure.
09:49And the agent suggests remove them, create a bouquet blur,
09:54maybe add a romantic sky, and a dramatic effect.
09:59You are still the creative director, right?
10:02It's your vision.
10:03But as soon as you decide on choice,
10:07the agent instantly executes.
10:11The last example of an agent that we're working on
10:14is an Adobe Express.
10:16Adobe Express is made as a lighter weight tool
10:19for those who may not have very deep creative training,
10:22like myself.
10:23And it works across video, audio, images,
10:28and it does translation as well.
10:30With this kind of agent,
10:31you don't actually have to understand
10:33the taxonomy of creative tooling.
10:36You can just tell it.
10:37I'm looking for this feeling or this overall look.
10:40In this example, more tropical.
10:43And then you iterate.
10:45Perhaps you don't like the first one.
10:47You choose.
10:47You direct.
10:48You give feedback.
10:49You keep coaching the agent
10:51until you get a product that you're happy with.
10:57Okay.
10:58I don't like to just show what's happening.
11:00I'd like to give you some implications.
11:02And one implication which may be a little counterintuitive
11:06is that human taste and ingenuity in the age of ideas,
11:11in the age of AI, becomes even more paramount.
11:15Why is that?
11:16Because the technical ability required to produce creative work,
11:21that bar is, that floor is lowered.
11:24And that's a great thing.
11:26It's a great thing for access across more communities.
11:29It's great for our young people,
11:32aspiring young creative pros.
11:34So the technical on-ramp gets lower.
11:38What that means is that what will differentiate creative products?
11:42It's you.
11:43It's your creative vision.
11:45It's your creative thinking and your imagination.
11:50Okay.
11:51Let's turn to the second unlimited possibility.
11:55Marketing.
11:56Personalized marketing at scale is something everyone wants to achieve.
12:01I work with a lot of brands who are trying to break through the noise
12:05of all that social content that we're always being bombarded with.
12:09And the old definition of personalization was,
12:13I, as the brand, produce a lot of content and I push it at you.
12:18The new definition of personalization is,
12:21I empower the consumer with the brand tools via AI
12:26and let them become the protagonist in my brand's narrative.
12:31Actually, they want to co-write that narrative with the brand
12:34and create something cool.
12:36And that presents a tremendous opportunity
12:40for the brand to engage very deeply with the consumer.
12:46So if anybody's been in digital marketing for any amount of time,
12:50we have been talking about personalization for a long time.
12:54And we've probably spent the last decade on the MarTech stack, right?
12:59The marketing technology aspect.
13:01Those are the two right things.
13:03Where's your data?
13:04Do you have a CDP?
13:05Do you have a dam to store assets?
13:07Do you have a journey optimizer?
13:08Do you have analytics for A-B testing?
13:10Great. We put all that in place in the last decade.
13:14But what did we find once we laid those pipes?
13:18We need more content.
13:20So content is what fuels growth.
13:22Engaging on-brand content fuels growth.
13:26And I don't know about you, but our CFO said you get no more budget.
13:31You're in a flat budget situation, but I need you to do more.
13:35The only answer to that is smart use of AI.
13:42Okay. So I like to do this.
13:44Again, I'm a math person.
13:45This is what I call the math problem.
13:47It's combinatorics.
13:49You have eight products.
13:51Imagine you only had eight products.
13:5215 channels, 35 languages, variants.
13:55You get the problem.
13:56If you only did 12 refreshes a year.
14:00We know the speed of social.
14:02We need to be doing three refreshes a week or that asset is dead on arrival.
14:07So just keep doing the math.
14:09You're going to reach very, very large numbers.
14:11So our answer to this usually is, I'm only going to support the top three products.
14:17I'm only going to support the top three markets.
14:20That does not allow us to grow.
14:24Further compounding this problem, well, most CMOs have said, I've been asked for 5x the growth in content.
14:34My creative team is spending less than a third of their time on actual creative work.
14:40And shocker, less than 10% of my approved assets in the dam are ever reused.
14:47That's, talk about sustainability.
14:50That's a huge waste problem that we have.
14:54Okay, so what lots of brands are doing now, and we're working with many of those top global brands,
15:01is thinking about the use cases.
15:03These are four that are no brainers to do content production at scale.
15:09One, localize your assets.
15:11You shoot it for a French speaking market or an English speaking market, you run out of money.
15:16Now, with smart AI, with responsible AI tools from Adobe, you can take that into 13 languages, you can translate
15:25everything, it can stay on brand.
15:27Refreshing the campaign seasonally, a lot of you shoot everything on white background because that's the easiest.
15:33But we know that things in situ convert much better in commerce.
15:38Now, you can actually do that for each season.
15:41Lower left, personalize.
15:43Again, back to N equals 1.
15:45We would love to talk to every customer as an N equals 1 segment.
15:49That was cost prohibitive.
15:51Now, if you hook this up with your CDP, you can get to a lot more segments and have the
15:56creative that supports it.
15:58And the last one, which I'll show in the lower right, is AI can actually create a whole innovative set
16:04of experiences that your customers can be immersed into.
16:09Okay, this is the one I was telling you about, Gatorade.
16:12So, we worked with Gatorade, owned by Pepsi.
16:16They call this Gatorade bottle the badge of the athlete.
16:20If you have any athletes in your family, I have two swimmers, the entire pool deck is lined up with
16:25these Gatorade bottles.
16:27Before Christmas last year, they came to us and said, we want to launch an AI bottle.
16:31We want the consumer, the athlete, to drive their vision of what the brand is.
16:37You want a bunny in a pool?
16:39You want an astronaut jogging?
16:41You can have it.
16:42And what you can see is it still respects the bolt, right?
16:46That's iconic.
16:47It respects the style that has gone with, but it still gives the tools to the consumer.
16:53This product outperformed their wildest expectations for Christmas.
16:57They are rolling out a whole series of additional products this year.
17:04Okay, another thing that you need to think about is video.
17:09Video converts nine times better than static, but nobody shoots video for all the formats.
17:16This is a bad crop, right?
17:18The bad crop is I lost the action, I lost the actor when I cropped.
17:22Here with Adobe's technology, we can take one hero shot, intelligently cropped to keep the actor and the product in
17:31view, and do this in a manner of seconds.
17:35This used to take people, you know, you're re-keying frame by frame, it's hours of work.
17:42And then I talked about...
17:44Let's see.
17:46We should have...
17:47There are a few key steps to take to ensure a safe and amazing adventure.
17:51When ever you plan in a state or national park to travel or travel, there are a few key steps
17:57that you need to take to ensure a safe and enjoyable adventure.
18:16Okay, so that's just a clip sample to tell you you can reach all the markets. This is an incredible
18:24technology we have. It's responsible, which means that we had rights to train all the voices that went into the
18:30model.
18:31And what you can do is shoot your hero actor once and take them into, at least right now, we
18:36have 13 languages.
18:38What you can see is the voice print stays the same, which means the audio of the actor remains through.
18:44And secondly, the lip action, right? The German versus the French.
18:48It's a little laggy here, but it's exactly synced in with the voice.
18:52So it opens up a whole new world of possibilities.
18:55Okay, at the end of this section on our second opportunity of personalized marketing, I want to tell you our
19:01own story because I lived through this story.
19:03Our Black Friday last year.
19:07We are a product company.
19:08Our CEO, Shantanu Narayan, likes us to be what we call customer zero, which is the first customer of every
19:15product.
19:16I call it drinking the champagne first.
19:19And we wanted to do exactly what I showed you, reach more markets.
19:23We do 30 languages, 16,000 assets.
19:26And this used to take us over 16 weeks to execute.
19:30In that process, previously, we would intentionally create creative waste,
19:36which means that we didn't know the price points that we would go out at at Black Friday.
19:41So we would produce every discount point you could possibly imagine in case we needed it.
19:46In this case, we were able to shrink that time by 60%, which brings us closer to the point of
19:52Black Friday giving us agility.
19:54And we freed up a lot of people's time so they could work on other really important campaigns.
20:00Huge success.
20:01It went all the way through from Adobe's DAM, Workfront, Frame.io for review and approval.
20:08And you can see some samples of the statics that we had here.
20:14So what's the implication of being able to do really great personalized marketing at scale?
20:21First, every brand will be able to deliver lots of personalized digital experiences for each one of us.
20:30And they'll be high quality.
20:32The second thing is the best brands will be producing 10x the content.
20:38But it'll be high quality content and it'll be on-brand content.
20:42What we don't need, and I saw the previous Cantor session, we don't need more bad creative.
20:47That's already happening.
20:48So we don't want AI to accelerate that.
20:51We want people to do better, high quality, on-brand content.
20:56And this one, I love this.
20:57It's liberating.
20:59Creative professionals will finally spend radically more time actually creating.
21:04So let's remove what they were bogged down with, right?
21:07All the drudgery, all those resizing pixel-by-pixel tasks,
21:11and let them do the creative vision and work that they were meant to do and brings them joy.
21:18They can use AI as a search engine for their imagination.
21:24The last idea, we've already covered the first two, right?
21:27Of unlimited possibilities.
21:29The first one was creative agents.
21:31The second one was personalized marketing.
21:33The last one I want to share with you is narrative and world building.
21:37This one is truly unlimited in its potential.
21:41Brands and creatives really need a complex, interconnected system,
21:46or universe that understands the brand.
21:50Or if you're making shorts or long-form, understands the franchise,
21:54the characters, the plot lines.
21:56To do that, they need hundreds of thousands of connected assets,
22:01and they also need expertise in AI.
22:05Why doesn't an off-the-shelf model work?
22:08One, it doesn't deeply understand the IP.
22:13Two, it can't maintain an environment that's cohesive.
22:18And three, therefore, it can't be consistent.
22:21In other words, off-the-shelf models, they lose the plot very easily.
22:26So, to create a solution to this, Adobe has brought to market Firefly Foundry,
22:32which is a commercially safe, deep-tuning model for world building.
22:39Before I get too far into what Foundry is,
22:42let me just explain how we think about customization.
22:45If you've used Firefly-based models, which are available right now,
22:51what you'll realize is that it doesn't know your IP.
22:54That's a feature, not a bug.
22:56If you try to put in Spider-Man, it has no idea what you're talking about,
23:01and that's part of our promise of commercial safety.
23:04Every single asset that went into training our models
23:07was a licensed piece of asset that a creator gave us permission to use,
23:11and they were compensated in the process.
23:13You can give this a single-shot reference, a sketch, another asset,
23:20and we'll do some light customization.
23:21The second thing we have in market is Firefly custom models for enterprises.
23:27These are light adapters.
23:29They sit on top of our base models.
23:32They never mix your content back with our base models,
23:35and you can train a single style, a single character, a single concept.
23:39So, this is great for fast work.
23:41This is great if you're doing like a campaign look.
23:44You can train on that style,
23:46but it only understands one concept at a time.
23:49What I was discussing around Foundry is world building.
23:53This is where we deeply tune our commercially safe models with your IP.
24:00That IP is yours only,
24:02and you're able to build an entire cohesive world.
24:10So, this is a little visual B-roll that we put together.
24:13We open up the base models which are commercially safe.
24:18We take in franchise IP in whatever form.
24:223D assets, video, image, audio, voice,
24:27and we retrain the models, billions of parameters,
24:32so that it understands your world.
24:33What you get at the end is a model that is tuned specifically for that franchise,
24:38and only owned by that IP owner.
24:41IP protection is obviously very important to Adobe,
24:44being at the center of the creative community.
24:46It never goes back into the base models,
24:48and that IP is yours.
24:51So, what can Foundry create?
24:54What you saw in the last, and here are the visuals for.
24:57We trained on an open source movie project.
25:00It's a blender studio project,
25:03and we trained a model on a lot of the characters.
25:06So, on the left, you can see output of the trained model.
25:10It keeps the plot.
25:12It understands the characters, the look, the creative direction,
25:15and this speeds up everything from storyboarding and ideation
25:19to final pixel-ready YouTube shorts.
25:25The other thing you can do with Foundry is hyperlocal models.
25:29So, this is an example of a Foundry Japan model that we trained.
25:34In the left, you can see the prompt.
25:36We want kids going to school.
25:38What it shows you are supposedly Japanese kids
25:41wearing a Western backpack.
25:44On the right, with the Foundry model,
25:46when we use that same prompt,
25:48it understands the concept of the randusoro,
25:52which is the special backpack that all the kids in Japan use.
25:56The same thing on the bottom, business cards.
25:58We do it very casually.
26:00I actually don't even know if we carry business cards anymore.
26:03But the Japan Foundry model understands that this is a formal, respectful exchange
26:09and uses the two-hand gesture for transfer.
26:12So, localizing Foundry models can be really great for culture-aware marketing.
26:19Okay, another part of Foundry, and because I'm a total geek,
26:23I just had to show you some of this,
26:25is bringing new research to market.
26:27So, this is text to 3D.
26:30And if anybody works in the 3D space,
26:32you'll understand how long it takes you to generate a full rendered 3D model
26:38with this amount of detail.
26:39This is done with a text prompt.
26:41I'm not even showing you the other research we have around Gaussian splat,
26:46where you can take a video capture and get a rendered 3D model.
26:50But this is incredible technology that we're also bringing to market as part of Foundry.
26:56Okay, the next one, we call this sneak vector turntable.
27:01If you're a vector artist, you'll love this.
27:05You can take a simple 2D,
27:07and the original artist only drew the front of the 2D image.
27:11Then the AI trained model is able to give you a complete 360
27:16in any direction render of that vector drawing.
27:22This is going to unlock a massive amount of time for folks.
27:27You probably have hundreds of sketches in your book.
27:29You can now actually develop them all.
27:34So, what are the implications for that last unlimited possibility of world building?
27:40Well, first of all, if you have Foundry models,
27:44time from ideation to final product shrinks by 10x.
27:49That's equivalent to defying time and space.
27:53And almost more importantly, ideas to the power of AI flourish.
28:00AI, my premise is it's going to remove the classic barriers and limits of time, money and resources.
28:09There are so many stories that aren't being told because there's a limit on how much budget they can raise.
28:17And so now, by lowering that ability to enter, we will hear so many more creative voices,
28:23and so much more creativity will be added into the classic canon.
28:28Well, I hope this session was helpful in some small way,
28:34and that you'll leave here looking for your version of unlimited possibilities.
28:40In closing, I always like to get very practical.
28:43How do I get started? What do I do tomorrow?
28:46Let me just give you three tips to leave with.
28:49First of all, don't try to change the entire world.
28:53Stay focused.
28:55Be small and intentional to start.
28:57Find one use case.
28:59Prove it. Scale it.
29:01Right?
29:02Fast, fail.
29:03If it doesn't work, move on.
29:04If you find a green shoot, scale it.
29:07The second thing is we need change agents.
29:11We need sponsorship and people who raise their hand
29:14and lean into the ambiguity of new innovation.
29:18So I would say lead from the front, tell the stories,
29:22and encourage your teams to embrace change.
29:26It's okay to learn, right?
29:27And it's okay to try and fail.
29:29And the last thing is I would say think about the nutrition label behind the models.
29:35We say nutrition label because you've got to think about
29:38what went into that model to create it.
29:40We have a particular creative approach to commercial safety,
29:45which we feel very firm about.
29:47But depending on your use case, just think,
29:50is how this model developed good for the use case that I'm in?
29:55All right.
29:56Well, thank you very much.
29:58I think...
29:59I appreciate it.
30:00Thank you.
30:05Have a great Vivitech and thanks for listening.
30:10Thank you.
Commentaires