Latest / AI Fire Daily / #6 Eric: Google’s Creative AI Ecosystem — Designing Brand Consistency Across Photo, Web, and Video

Transcript

#6 Eric: Google’s Creative AI Ecosystem — Designing Brand Consistency Across Photo, Web, and Video · AI Fire Daily

0:00
Think about the ultimate frustration with AI
0:02
right now. Generating one stunning image is incredibly
0:06
easy, but getting the machine to recreate that
0:09
exact style, it is an absolute nightmare. Well,
0:12
yeah. You tweak one single word, suddenly the
0:15
whole aesthetic just completely falls apart.
0:17
Right. It's just a mess. Beat. So, welcome back
0:20
to the Deep Dive. Today we're unpacking the TLDR
0:23
BIOCX newsletter. We're looking at Google's creative
0:26
AI ecosystem. And our mission today is, well,
0:30
super practical. We are going to build a brand
0:32
from scratch. Yeah, we're building a fictional
0:34
premium menswear brand today. We're calling it
0:37
Valen. We're taking it from just an idea to a
0:40
fully functioning campaign. Exactly. And we'll
0:42
move logically through four specific tools. We
0:45
use Gemini, Pimeli, Stitch, and Flow. The goal
0:48
here is to formally prove a critical insight.
0:52
It is never simply remembered by the machine.
0:55
I absolutely love this experiment. The biggest
0:56
shift in modern AI isn't better generation. It's
0:59
the unprecedented ability to carry approved work
1:02
forward. You don't have to start from zero every
1:04
single time. Right. You build a foundation and
1:07
you carry it with you. But before we can build
1:08
a brand, we really must lock down reality. We
1:12
have to stop the AI from hallucinating product
1:15
details. That's huge. So let's formally introduce
1:18
our hero product for this. We are calling it
1:20
the Regent Suit. We need to be very specific
1:23
about what this is. It's a deep charcoal two
1:26
-piece men's suit. It features a beautifully
1:29
tailored fit and a classic notch lapel. And it
1:33
uses a premium wool blend construction. See?
1:36
That is a highly specific physical product. You
1:38
cannot afford for the AI to change those details.
1:41
No, absolutely not. So Gemini is the flexible
1:44
starting point for this raw brief. We use it
1:46
to establish our fundamental ground rules. Yeah,
1:49
and that very first step is absolutely crucial.
1:51
You must strictly prompt Gemini to separate your
1:53
input into two distinct groups. Group one is
1:56
called your locked facts. These are the fundamental
1:59
things the AI simply cannot invent. Think about
2:01
the physical wool material, or your target audience.
2:04
And then group two is your flexible creative
2:06
freedom. The elements we actually want the AI
2:08
to play with. Precisely. The creative freedom
2:11
covers the emotional mood of the campaign. It
2:14
dictates the setting and the underlying photography
2:16
style. You're building a structural boundary
2:19
for the machine. The locked facts act as a rigid
2:22
structural guardrail. You are actively telling
2:25
Gemini what is strictly off limits. Because the
2:28
AI desperately wants to be helpful. It's highly
2:31
creative. It's basically a digital people pleaser.
2:34
Right, it really is. Without those guardrails,
2:36
it might spontaneously decide your suit needs
2:39
a silk lining. Or it might randomly invent a
2:43
fictional history for your brand. That is a massive
2:46
legal nightmare for real businesses. You cannot
2:48
advertise features your product does not actually
2:51
possess. Exactly. That separation gives the software
2:54
safe room to breathe. It can beautifully explore
2:56
dramatic lighting safely, but it won't mess with
2:59
the actual physical product specifications. It
3:02
knows the difference between an aesthetic and
3:03
a fact. Hold on though. I want to firmly push
3:06
back on that just a bit. Let's say I set the
3:08
creative direction to Italian inspired. Isn't
3:11
there a huge risk the AI rewrites my product
3:14
specs? Like, suddenly claims the suit is actually
3:17
handmade in Milan. Those linguistic boundaries
3:20
seem incredibly fragile. I still wrestle with
3:23
prompt drift myself. Oh, that is a very real
3:25
danger in this process. And for those listening,
3:28
prompt drift is common. It's when the AI slowly
3:30
forgets your original instructions over time.
3:32
Yeah, and that is exactly why we lock the facts
3:35
down first. If you just ask for an Italian -inspired
3:38
suit campaign, the model guesses, it naturally
3:40
assumes Italian means manufactured in Italy.
3:43
By separating the locked facts, Gemini acts as
3:46
a filter. It holds the line against those linguistic
3:49
assumptions. Pete, so why do we use Gemini first
3:52
instead of jumping right into an image generator?
3:54
Well, image generators do not understand logic
3:56
or brand safety. They just desperately want to
3:59
make the picture look visually stunning. Gemini
4:02
acts as a logical reasoning engine behind the
4:04
scenes. It mathematically organizes your constraints
4:07
first. This ensures the image generator stays
4:10
completely honest to the brief. So Gemini acts
4:12
as the brain organizing facts before the creative
4:15
muscles flex. Right. And it saves you hours of
4:17
frustrating rework later. OK. So our core facts
4:21
are locked. and completely accurate now. But
4:25
a list of facts is not a fashion brand. We need
4:28
to seamlessly translate them into a visual identity.
4:31
It needs to look like a real functioning business.
4:33
Yeah, exactly. This is where we transition to
4:35
a tool called Pameli. Pameli is fascinating because
4:38
of its core philosophy. It focuses entirely on
4:41
building business DNA. It doesn't care about
4:44
processing isolated one -off image prompts. It
4:47
wants to know the soul of your company. Right.
4:49
And that is how Vellan truly takes its final
4:51
shape. We have to establish the primary guiding
4:54
principle for the brand. For Valen, that principle
4:57
is soft confidence over hard power. We give the
5:00
Pameli software a heavily locked visual system.
5:03
We establish the official architectural twilight
5:05
color palette. We're talking deep slate blues
5:08
and like sharp concrete grays. You also lock
5:11
down the typography for the brand. We chose the
5:14
Bodonimoda and Jost fonts and we established
5:17
the official brand tagline. It's Vanelline, the
5:21
quiet standard. But getting this right requires
5:23
a highly repetitive iterative process. I mean,
5:26
the first few campaign images were just not good.
5:29
They were way too text heavy and cluttered. Yeah,
5:32
they looked like generic department store mailers.
5:35
They completely lacked that premium, quiet, luxury
5:38
feeling we wanted. Exactly. That iteration loop
5:41
is completely normal in this specific workflow.
5:43
You have to widely explore before you finally
5:45
lock it down. Pameli might occasionally return
5:48
weird collages instead of single clean images.
5:51
Yeah. You cannot expect perfection on the very
5:54
first click. You just curate the strongest visual
5:56
assets and confidently move forward. The human
5:58
art direction is still what makes the output
6:00
good. We had to really refine the text prompt
6:03
to fix the cluttered look. The new prompt was
6:05
very direct and structural. Minimal luxury fashion
6:08
campaign. Photography first. Generous negative
6:11
space. No text over faces. And that instantly
6:14
solved the aesthetic problem we were having.
6:16
It really did. And once you get that perfect
6:18
image, you save it. That image becomes the visual
6:22
anchor for the entire brand context. It gives
6:24
the AI a highly reliable starting system. It
6:28
aggressively stops the software from wandering
6:30
off in the direction. Bede, if the first image
6:33
is great, why is it dangerous to just write a
6:35
new text prompt for the next one? A brand new
6:38
text prompt forces the AI to reinterpret everything.
6:41
It basically reimagines the entire brand visual
6:44
from scratch. It reads luxury suit and just guesses
6:47
at the lighting again. Which is bad. Yeah, that
6:49
leads straight to immediate and frustrating visual
6:51
drift. Using your approved images strongly anchors
6:55
the AI to your specific design. Right. Text invites
6:57
the AI to guess, but reference images lock the
7:00
design down. Yeah, that is the most accurate
7:02
way to frame it. You want the machine referencing
7:04
pixels, not interpreting adjectives. So we have
7:07
gorgeous, highly accurate brand imagery right
7:10
now. The suit looks real. The lighting looks
7:12
expensive. But a fashion brand cannot just exist
7:15
as a folder of JPEGs. Customers need a place
7:18
to actually buy this suit. Right. A brand has
7:20
to survive a user interface to be functional.
7:23
Which brings us to the next massive hurdle. Let
7:26
me introduce Stitch to the creative workflow.
7:29
Stitch is Google's dedicated tool for UI and
7:32
digital design. It is built to generate web layouts
7:35
and application screens. And there is a massive
7:38
operational trap waiting right here. People generally
7:42
start by generating a beautiful homepage design.
7:44
They get the fonts and the spacing just right.
7:47
Then they inevitably make a huge mistake. They
7:50
naively generate the product page from a completely
7:53
fresh text prompt. And the digital visual language
7:55
immediately starts to aggressively drift. The
7:58
company logo suddenly moves from the top left
8:00
to the center. Right. The structural spacing
8:02
changes completely between the two pages. The
8:05
navigation bar looks like it belongs to a different
8:07
company entirely. The individual digital screens
8:09
might actually look fine on their own, but they
8:12
absolutely do not look like they belong together.
8:15
Aesthetic variations quickly pile up on the digital
8:17
canvas. Important details like typography and
8:20
layout start to completely fall apart. It's the
8:22
exact same problem we had with the images earlier.
8:25
You are asking the AI to remember a vibe. AI
8:28
does not remember vibes. It just generates based
8:30
on the current prompt. So what is the highly
8:32
practical software solution here? You generate
8:36
the initial home page and officially approve
8:38
it. Then you explicitly use that specific screen
8:41
as the master reference. Yes. It securely anchors
8:44
the collection and Atelier pages directly. You
8:47
are forcing the AI to look at the approved layout.
8:50
You specifically point the AI back to the approved
8:52
output. You aggressively stop manually describing
8:56
the visual style from memory. You just say, make
8:58
the product page look exactly like this homepage.
9:01
Exactly. Building UI this way is like stacking
9:03
Lego blocks of data. You aren't describing the
9:06
block every time. Each new page snaps perfectly
9:08
onto the digital foundation of the last. The
9:11
structural integrity remains completely intact
9:13
across the whole site. Beat. Does having a polished
9:18
stitch design mean the website is fully ready
9:20
to launch? Not at all. Stitch strictly makes
9:23
high fidelity visual prototypes. It shows you
9:26
exactly what the website will look like. But
9:28
it does not build functional backend logic or
9:31
payment systems. It is just the beautifully designed
9:33
visual interface layer. Got it. Stitch builds
9:36
a beautiful prototype, but you still must code
9:38
the backend. Yes. It purely handles the creative
9:42
front -end layout direction. The engineering
9:45
still requires actual developers. So static websites
9:48
and still images are mostly conquered at this
9:51
point. We have a locked product. We have a visual
9:53
identity. We have a beautiful prototype website.
9:57
The ultimate test of visual consistency is making
9:59
the brand actually move. We have to expertly
10:01
generate video without reality visually melting.
10:05
Yeah. Motion is where generative AI traditionally
10:07
falls completely apart. Time adds millions of
10:10
new pixels into the equation. Let's clearly differentiate
10:13
the two main video tools first. Google Vids is
10:16
for highly structured, editable business videos.
10:19
Think about standard corporate explainers or
10:22
internal training presentations. Right. It handles
10:24
bullet points and pie charts beautifully. But
10:27
we are building a premium fashion brand. We need
10:31
highly cinematic media instead. Enter Flow. Flow
10:34
is purposefully built for complex cinematic video
10:37
projects. It handles narrative storytelling and
10:40
high -end commercial aesthetics. You use Flow
10:43
Agent to carefully plan the commercial storyboard.
10:45
And the storyboard for Valen is simple but highly
10:48
effective. It moves through four distinct beats.
10:51
Atmosphere. Discovery. Putting on the jacket.
10:54
And finally... The hero shot. The crucial step
10:57
here is bringing in the approved Pimeli assets.
11:00
We take those perfect campaign photos and feed
11:02
them into Flow. You firmly lock the character
11:05
in the video as Valen Man. This ensures the visual
11:07
identity remains perfectly consistent across
11:09
different shots. Exactly. Use a matching full
11:12
body reference image for the software. Flow now
11:15
has one beautifully consistent identity to reuse
11:17
throughout the commercial. The actor doesn't
11:19
magically change between scenes. Whoa! Imagine
11:22
scaling to a billion queries, generating personalized
11:25
cinematic video campaigns in real time, to sex
11:28
silence. That is absolutely incredible to think
11:30
about. The creative possibilities are genuinely
11:32
staggering. It is undeniably wild. But you have
11:36
to tightly control it using master frames. These
11:39
are still images used to tightly define the visual
11:41
composition. You specifically do this before
11:43
adding motion via the VO3 .1 fast model. And
11:47
VO3 .1 fast model is Google's high -speed AI
11:50
engine for generating realistic video clips.
11:54
So you meticulously get the still image right
11:56
first. You check the lighting. You check the
11:58
suit details, then you slowly introduce the complex
12:01
motion afterward. Right. And for complex human
12:04
motion, you need specific techniques. Think about
12:06
the scene of putting on the jacket. That is incredibly
12:09
hard for AI to render accurately. The arms often
12:13
morph through the fabric. Oh, yeah. The lapels
12:15
magically change shape mid -motion. The geometry
12:18
of the suit just completely breaks down. To fix
12:21
this, you use the first frame to last frame technique.
12:24
You give the video model incredibly clear starting
12:26
and finishing states, you show it the man holding
12:28
the jacket, then you show it the man wearing
12:30
the jacket. Wow. The AI smoothly connects the
12:33
states instead of inventing the action. You are
12:36
drastically limiting the variables it has to
12:38
calculate. We do need to be honest about where
12:40
the AI still falls short, though. Flow can still
12:44
suffer from highly noticeable visual morphing.
12:47
Faces, jackets, and room details can start shifting
12:50
mid -shot. Yeah, definitely. The background environment
12:53
details can get a little weird sometimes. Books
12:55
on a shelf might just like melt into each other.
12:57
Yeah, the dressing motion usually requires a
13:00
few different attempts. It is rarely perfect
13:02
on the very first render. The main fix is carefully
13:05
keeping your video shots simple. Do not ask for
13:08
wild camera pans or complex blocking. You must
13:12
firmly anchor the generation with those approved
13:14
visual references. Right. And good -looking output
13:17
is definitely not the same as being on brief.
13:19
The video might look incredibly cinematic and
13:22
beautiful, but if the suit suddenly has three
13:24
buttons instead of two, it is useless. And generative
13:27
edits can also seriously disturb correct details
13:29
in an image. You try to fix one small part of
13:32
the frame, suddenly you break another part that
13:34
was already perfectly right. That happens all
13:36
the time in post -production. You have to make
13:38
highly focused, localized edits. Always save
13:41
strong versions so you can easily revert when
13:43
needed. Yeah. Beautiful output can still be completely
13:47
wrong logically. The AI can invent believable
13:51
product facts while still looking highly polished.
13:54
You have to relentlessly review everything against
13:56
the original brief. You cannot just judge it
13:58
based on the superficial visual quality. And
14:00
the so -called best model may not be the right
14:02
one either. Different models strongly support
14:05
different reference and continuity controls.
14:08
You have to choose based on the specific control
14:10
you need first. Right. If maintaining the lapel
14:13
structure is important, use a stronger model.
14:15
Do not desperately try to rescue a drishing shot
14:18
later. Fix it at the source. Beat. Why create
14:22
still master frames first instead of just prompting
14:24
the AI for a video directly? Still frames securely
14:27
let you perfect the lighting, character, and
14:29
environment. You tightly lock all of that down
14:32
before introducing the complexity of motion.
14:34
Motion simply adds way too many unpredictable
14:37
variables at once. It overwhelms the AI's ability
14:40
to maintain consistency. Stills lock the composition
14:43
and lighting before introducing the absolute
14:46
chaos of video motion. Exactly. You tightly control
14:49
the visual environment completely first. Then
14:52
you generate the brief motion clips. Finally,
14:55
you take the strongest clips into Scene Builder
14:58
to assemble them. Let's clearly synthesize this
15:00
entire creative journey today. We took a blank
15:02
page and built a functioning brand. The software
15:05
ecosystem is a highly structured creative pipeline.
15:08
You use Gemini for highly flexible logic and
15:11
planning. You establish the unchangeable rules
15:13
of reality. Then, Pamelli firmly locks in your
15:17
visual brand DNA. It turns those rules into a
15:21
recognizable aesthetic. Next, Stitch expertly
15:24
handles your digital interfaces and functional
15:26
prototypes. It builds the digital storefront
15:28
without breaking the visual rules. Finally, flow
15:30
and vids are strictly used for motion. They bring
15:33
the brand to life cinematically. The golden rule
15:35
of modern AI creation is incredibly simple. Explore
15:38
with AI. Approve with a human. Carry the approved
15:41
work forward. Never ask a new AI tool to reinvent
15:44
what the previous one already got right. You
15:47
logically just reuse the approved digital assets.
15:50
You efficiently carry the best previous output
15:52
forward instead of restarting. You are building
15:55
a chain of trust across the different software
15:58
platforms. Right. Brand consistency is actively
16:01
designed by the human operator. It is definitely
16:04
not just passively remembered by the machine.
16:06
The machine only knows what you show it in that
16:09
exact moment. Here is your practical homework
16:11
from the newsletter today. Pick a real or fictional
16:14
product this week. Strictly separate the locked
16:17
facts from the creative freedom using an LLM.
16:20
Explore several directions and strongly approve
16:22
one visual identity. Yeah, that's a great exercise.
16:26
Then carry it into at least two different outputs
16:28
to test the consistency. Try making an image
16:30
and then a website mockup. You want to clearly
16:33
see if they still belong together. Carefully
16:35
note where the visual identity grifted during
16:37
the entire process. Did the color shift? Did
16:40
the mood change? It is a profound exercise in
16:43
understanding how these tools actually think.
16:45
And it brings up a massive question for the future.
16:49
If AI can eventually maintain a brand's visual
16:52
identity perfectly across every possible medium,
16:55
wait, I shouldn't trail off. If AI can eventually
16:57
maintain a brand's visual identity perfectly
16:59
across every possible medium, will the future
17:02
of creative work be less about making the art
17:04
itself and entirely about designing the constraints
17:06
that govern the machine? That is a profoundly
17:09
fascinating question to leave you with today.
17:11
We have to end the deep dive on that note. The
17:13
role of the creator is fundamentally shifting
17:14
before our eyes. Thank you for deeply exploring
17:17
this complex topic with us today. We will absolutely
17:20
see you next time.