Latest / AI Fire Daily / #567 Neil: Create Realistic AI Videos With Claude ChatGPT And Flow

Transcript

#567 Neil: Create Realistic AI Videos With Claude ChatGPT And Flow · AI Fire Daily

0:00
We have all been there. You drop actual money
0:02
on credits for the newest, shiniest AI video
0:06
platform. You wait five agonizing minutes for
0:09
it to render. And congratulations. You get a
0:12
guy with seven fingers. Exactly. You have just
0:15
paid for a video of a guy with a melted face.
0:18
It is incredibly frustrating. It really is. I
0:20
mean, it feels less like creating and more like
0:22
pulling the lever on a very expensive slot machine.
0:25
Welcome to the deep dive. We are skipping the
0:28
hype cycle today. It feels like a new AI video
0:31
platform drops every single week right now. Oh,
0:34
at least once a week. Right, and everyone promises
0:37
perfectly realistic motion. But the results,
0:40
they usually look unfinished. They just, you
0:41
know, they look obviously generated. Yeah, and
0:43
constantly testing those new platforms costs
0:45
real money. Every failed attempt just burns through
0:48
your monthly credits. It is expensive. Totally.
0:51
That trade off simply does not make much sense
0:52
anymore. So today we are looking at something
0:55
completely different. We are exploring a workflow
0:58
that is a bit older, but it is pure gold. It
1:01
prioritizes absolute meticulous control over
1:04
random generation. It's a highly deliberate process,
1:08
very intentional. We're going to walk you through
1:10
a very specific three tool stack today. We will
1:13
use Claude for visual style. We will use ChatGPT
1:17
5 .6 for our reference images. And finally, Google
1:21
Flow for the actual video generation. I know
1:23
it sounds like a lot of jumping around, but once
1:25
you understand the underlying logic, it just
1:28
changes everything. Let's unpack this blueprint.
1:31
Before you build a physical house, you need an
1:33
architectural blueprint. So if our goal is to
1:36
make a highly realistic video, Why are we starting
1:39
with a text -based AI? Good question. Why not
1:41
just jump straight into an image generator? Because
1:44
vague prompts create messy videos. It is literally
1:47
that simple. Right. If you just open a video
1:50
generator and ask for a cool cinematic shot,
1:53
the AI has to guess. It fills in the blanks.
1:56
And frankly, AI is terrible at guessing. So you've
1:58
got to choose a specific capture style first.
2:00
Exactly. This choice shapes everything else down
2:02
the line. It defines the lighting. It defines
2:04
the camera angle. It dictates the color grading.
2:07
It establishes the overall baseline for realism.
2:10
Spot on. You have to choose one clear capture
2:12
method. You have to think about the exact camera
2:15
setup you want to simulate. Are you pretending
2:18
to be a fan recording on a smartphone from the
2:21
bleachers? Right. Is it a documentary crew with
2:24
a shaky handheld rig? Yeah. Or a fixed cinema
2:27
camera on a heavy tripod? You have to make that
2:29
decision immediately. When you describe the capture
2:31
style clearly, the machines understand the physics
2:34
of the scene. Chat GPT 5 .6 and Google Flow can
2:38
produce a much more mathematically consistent
2:40
result when they have strict parameters. I have
2:42
to admit a vulnerability here. I still wrestle
2:45
with prompt drift myself. Oh sure. I start with
2:48
a gritty sci -fi scene and end up with a cartoon.
2:51
It just slips away from me. That happens to everyone.
2:54
Well, it's a known vulnerability in how these
2:56
neural networks operate. They want to drift toward
2:58
average generic concept. They regress to the
3:00
mean. Exactly. That is exactly why we use Claude
3:04
Opus 5 for this very first step. Why Opus 5 specifically?
3:08
Because Anthropic designed Opus 5 specifically
3:11
for complex reasoning. Plus, it handles high
3:13
-level vision input beautifully. It can look
3:16
at an image and dissect the underlying technical
3:18
data. So the practical step here is that you
3:21
upload a reference image to Claude. You go online,
3:24
find a still frame from a movie or a photograph
3:27
with a visual style you love. but you probably
3:30
don't know how to describe the technical lighting
3:32
yourself. Right, and Opus 5 notices those tiny
3:35
details for you. It sees the nuances in the lighting
3:37
contrast. It feels the specific lens type. It
3:40
analyzes the color treatment. It gives you precise
3:44
technical data instead of a vague artsy summary.
3:47
So you feed Claude a highly specific prompt.
3:50
You ask it to analyze exactly how the original
3:52
footage was captured. You are asking for the
3:55
camera and lens feel, the lighting, the color
3:58
grading. The framing. And the level of realism.
4:01
But here is the absolute crucial instruction.
4:04
You must tell a law to give you one concise style
4:07
description. You must strictly command it to
4:09
avoid poetic or vague language. Does using non
4:12
-poetic language actually change how the AI interprets
4:16
the image data? Absolutely. Poetic language introduces
4:20
ambiguity into the latent space. If you say a
4:23
scene is bathed in a sea of golden tears, the
4:27
image generator might literally draw a sea of
4:29
crying eyes. Technical terms give the machine
4:32
exact mathematical boundaries for pixel generation.
4:36
So literal descriptions prevent the AI from hallucinating
4:39
unnecessary artistic flourishes. That is exactly
4:42
right. You take that dry literal description
4:45
from Claude and you save it. That text becomes
4:48
your anchor. It helps you keep the exact same
4:50
visual style throughout the entire rest of the
4:52
workflow. Exactly. So Claude has given us the
4:54
mathematical boundaries of our lighting and lens.
4:57
We have the blueprint, but a blueprint isn't
4:59
a house. Right. How do we start laying the physical
5:02
bricks without the AI hallucinating? This is
5:04
where we transition from text to our first foundational
5:06
image. We need to cast our actor and scout our
5:09
location. And for this you recommend using ChatGPT
5:11
5 .6? I do. Why move away from Claude at this
5:14
stage? Because ChatGPT 5 .6 is uniquely strong
5:18
at creating detailed persistent characters. It
5:21
handles realistic faces better than almost anything
5:23
else right now. It understands skin texture.
5:26
It understands how clothing folds. It gets the
5:29
micro details right. Why does that matter so
5:32
much? I mean, we're just going to animate it
5:34
later anyway. Because its image editing is incredibly
5:37
precise. Video generators like Google Flow will
5:40
use your still image as the absolute foundation
5:43
for movement. If your starting frame is flawed,
5:46
your video will be a nightmare. When your first
5:48
frame looks hyper realistic, your final video
5:51
looks far more stable. So the workflow here is
5:54
that you upload a clear photo of a character.
5:57
You want to preserve specific facial features.
5:59
Then you paste Claude's exact visual style description
6:02
right alongside it. Let's describe a specific
6:04
scenario for you. You are prompting ChatGPT to
6:07
use the uploaded person as your main character.
6:10
You explicitly tell it to preserve the facial
6:12
features, the hairstyle. the body proportions,
6:16
their core identity. Let's say you ask for a
6:18
16 by 9 realistic image. Our character is walking
6:22
alone on a mountain trail. It is late afternoon.
6:24
And you apply Claude's technical visual style.
6:27
You prompt for natural lighting, realistic skin
6:29
texture, believable shadows falling across the
6:32
trail. You ask for imperfect camera framing to
6:35
sell the realism. And you explicitly ban certain
6:39
things. You say, no CGI details, no text, no
6:42
extra people. So ChatGPT generates the image.
6:46
You have to check the face and hands carefully,
6:48
check the clothing, check how the light hits
6:50
the trees, but inevitably something will be slightly
6:54
off. Yes, and here is the golden rule of AI editing.
6:57
If one detail looks wrong, do not regenerate
7:00
the whole image. Ask ChatGPT 5 .6 to edit just
7:05
that specific part. Right, because if you regenerate,
7:08
you lose the magic of the parts that actually
7:10
worked. Exactly. You maintain your base image.
7:13
If the left hand is weird, you tell ChatGPT,
7:15
fix only the left hand. Once you have a strong,
7:18
flawless frame, you lock it in. It becomes the
7:20
main reference for your next shot. Yes. It's
7:22
exactly like stacking Lego blocks of data. Yes.
7:24
You lock the base piece down and build on top
7:27
of it. You don't rebuild the entire foundation
7:28
just to change one window. That is a brilliant
7:31
analogy. You build consistency block by block.
7:35
This step -by -step process helps ChatGPT preserve
7:39
the character, the outfit, and the exact location.
7:42
Let's say we want a second shot for our sequence.
7:44
We want to change the camera angle. Okay. We
7:46
tell ChatGPT to move the camera behind the person,
7:49
but now we want to show a grizzly bear farther
7:53
ahead on the trail. Ooh, dramatic. We tell it
7:56
to preserve the same time of day and realistic
7:58
camera quality from the first shot. you are creating
8:01
this new shot directly from your approved first
8:04
frame. You never start from scratch. Right. Because
8:07
you do it this way, Google Flow will eventually
8:09
receive a perfectly consistent set of images.
8:11
The lighting matches, the clothes match. Wait,
8:14
if I want a hiker and a bear, why not just ask
8:16
ChatGPT to generate both the hiker and the bear?
8:19
in the very first prompt. Because AI image generators
8:23
have a notoriously short attention span for spatial
8:26
relationships, if you ask for two complex subjects
8:29
immediately, it blends their traits. The hiker
8:31
might suddenly have fur. Introducing them sequentially
8:34
anchors the main subject first so the AI's neural
8:37
network doesn't get confused. Got it. Too many
8:40
subjects at once dilutes the AI's focus and consistency.
8:44
Precisely. You have to guide its focus one subject
8:46
at a time. Sponsor. This episode of the deep
8:49
dive is brought to you by our sponsors. We'll
8:51
be right back Okay, we are back. So we have a
8:54
beautiful mathematically consistent set of still
8:58
frames Now comes the most dangerous part of this
9:01
whole process. Yeah adding time and motion We
9:04
have to figure out how to add physics without
9:06
letting the AI warp our characters into Monsters,
9:10
this brings us to step three During those flat
9:13
reference images into moving video with Google
9:15
Flow. In step two, you successfully created two
9:18
still images with two different camera angles.
9:21
We have to turn each image into a separate video
9:23
clip. So we open Google Flow, we use the image
9:25
to video feature or frames as it's called on
9:28
their platform. They will join them together
9:29
later. Google Flow is built specifically for
9:31
this. You start with your first image. The character
9:33
is walking alone on the mountain trail. You select
9:35
the video tab. Then you choose frames. You add
9:38
the still image as your starting frame. You keep
9:41
the aspect ratio at 16 by 9. Then you have to
9:44
write an animation prompt. And the rule here
9:46
is to keep it very simple. Very simple. You just
9:48
say, the hiker walks slowly, clothing moves in
9:52
the wind, the camera follows with subtle handheld
9:55
movement. Keeping it simple is critical. When
9:58
you do that, Google Flow keeps the face and body
10:01
proportions consistent. The lighting and environment
10:04
match the reference image perfectly because it
10:06
isn't trying to calculate too many complex actions
10:08
at once. Now, we have to define the audio for
10:11
this clip. We prompt for soft footsteps, light
10:14
wind, distant birds. But importantly, we explicitly
10:18
say no music. Right. And you notice that animation
10:21
prompt only asked for one main action. The walking.
10:24
That constraint helps Google Flow keep the movement
10:27
much more natural. It doesn't have to calculate
10:29
a complex sequence of events. So clip one is
10:31
rendering. Then we animate clip two. This is
10:34
the image where the camera is behind the hiker
10:35
and the bear is ahead. We upload it. Our prompt
10:39
this time is just that the bear takes one slow
10:41
step forward. The hiker steps backwards. And
10:44
you prompt the camera to stay behind the hiker
10:46
with very light handheld motion. Because the
10:50
prompt is simple. Google Flow keeps everything
10:52
structurally sound again. Whoa. Beat. Imagine
10:57
the AI calculating the physics of every single
11:01
piece of gravel shifting underfoot. It has to
11:04
simulate light, shadow, and weight for thousands
11:07
of pebbles all at once. It is staggering when
11:10
you really stop to think about the math. The
11:12
sheer computational power required for that is
11:14
mind -bending. It really is. The machine is essentially
11:17
hallucinating physics in real time, frame by
11:20
frame. For the audio on clip 2, we prompt for
11:23
gravel movement. We want subtle bear sounds,
11:26
a bit of wind, but again, strictly no music.
11:28
Now you have both of your moving clips. You have
11:30
to combine them. You add them to a project scene.
11:32
This introduces something called Scene Builder.
11:34
Which is basically a digital editing timeline
11:36
where you arrange clips into one sequence. Exactly.
11:38
It looks like a simplified version of standard
11:40
video editing software. You place clip one on
11:42
the timeline first, then you place clip two right
11:44
after it. You hit play and you preview the full
11:47
sequence. You are looking to see if the transition
11:50
between the two shots feels natural. But because
11:52
we created these two clips completely separately,
11:55
the transition is probably going to look a bit
11:57
rough. The character's position or the movement
11:59
speed might change too suddenly at the cut. That
12:01
is incredibly common. When that happens, you
12:05
use a feature in Scene Builder called the Describe
12:07
Your Edits box. You actually ask Google Flow
12:10
to smooth the transition for you. You tell it
12:12
to create a natural visual bridge. So you tell
12:15
the AI to bridge the final two seconds of the
12:18
first clip and the opening two seconds of the
12:21
second clip. But you have to set boundaries.
12:23
Yes, strict ones. You strictly forbid sudden
12:26
cuts. You command it. No character morphing,
12:29
no dramatic sound effects. You are forcing the
12:31
machine to match the camera direction. It has
12:33
to match the subject position. It has to match
12:36
the movement speed. It even blends the environmental
12:39
audio. It smooths the connection without fundamentally
12:43
changing the main content of either clip. I noticed
12:45
you were incredibly specific about the audio
12:47
earlier. Why specifically ask it to not include
12:51
music in the individual clips? Because audio
12:53
generation in these platforms is baked directly
12:56
into the final video file. If you have baked
12:58
-in music on clip 1 and different baked -in music
13:01
on clip 2, you can't smoothly cut them together
13:04
without a jarring audio glitch. You want clean
13:07
stems, meaning isolated audio tracks completely
13:10
separate from any background music. Makes sense.
13:13
Embedded music restricts your editing choices
13:15
later in the final timeline. Exactly. Keep the
13:17
audio stems clean and isolated so you have total
13:20
control. So the clips are on the timeline. They're
13:22
bridged. But before we hit export, we have to
13:25
put on our director's hat. We enter the final
13:27
phase of this workflow. We have to rigorously
13:30
review and improve the results. This is where
13:33
most people fail. You have to do aggressive quality
13:35
control. AI video generators are infamous for
13:38
hiding terrifying flaws in the tiny details.
13:41
Yeah. A clip might look breathtaking at first
13:43
glance, but profound structural problems appear
13:46
when you look closely. Because the temporal smoothing
13:49
tricks our human eyes. So you must watch every
13:52
single clip twice. You watch once at normal speed
13:56
to check the vibe. Then you watch it once in
13:58
extreme slow motion. You are looking for very
14:01
specific failures. Does the face stay mathematically
14:04
consistent? Do the hands and body keep their
14:07
natural shape, or do they melt? Does the movement
14:10
look smooth and physically believable, or does
14:13
it look like they're floating? You check the
14:15
clothing for weird morphing. You check the lighting
14:18
to ensure shadows fall correctly. You make sure
14:20
the generated sound actually matches the physical
14:23
action on the screen. If a detail is wrong, you
14:26
do targeted fixing. This is crucial. You fix
14:28
one specific problem at a time. You describe
14:31
the exact problem in the prompt box. So they're
14:34
just complaining. You say, the character's right
14:35
hand looks distorted in frame 40. Exactly. Or
14:38
the walking motion looks too fast. Or the face
14:41
changes during the final two seconds. Do not
14:43
ever use broad prompts. Never just say, make
14:47
it better or fix this. A broad request causes
14:50
Google Flow to recalculate the entire vector
14:54
space. It will alter parts of the video that
14:56
were already perfect. It scrambles the good data
14:59
to try and guess how to fix the bad data. So
15:01
you tell it to keep the exact same character
15:04
and lighting. Fix only the specific issue you
15:06
named. You command it to make the corrected movement
15:09
look physically believable, and you explicitly
15:11
tell it not to change any of the part of the
15:13
clip. But here is the harsh truth of this workflow.
15:17
Do not keep a shot just because the AI finished
15:19
generating it. Right. It is physically painful
15:22
to delete work that took 10 minutes to render,
15:24
but it is true. A clip might have beautiful,
15:27
smooth camera movement. The lighting might be
15:29
breathtaking. But if the main character's face
15:32
morphs for even one single second? You have to
15:35
scrap it. You have to delete it and regenerate
15:37
it from the original reference image. You have
15:39
to be absolutely ruthless as a director. A single
15:42
morphing face destroys the illusion of reality
15:45
for the entire sequence. Your audience will instantly
15:49
know it's fake. You only add completely flawless
15:52
clips to your final video. Is there a psychological
15:54
trap here where we accept a bad clip just because
15:57
it took a long time to generate? Oh, absolutely.
16:00
It is a textbook example of classical conditioning.
16:03
You wait patiently for five minutes. You get
16:05
a visual reward. Even if the reward is deeply
16:07
flawed, your human brain wants to accept it just
16:11
to justify the waiting time. You have to actively
16:13
fight that instinct. Yeah, the sunk cost fallacy
16:16
is the enemy of a realistic video. Spot on. Don't
16:19
let your emotional sunk costs ruin your final
16:21
export. Delete the bad clips. Zooming out from
16:24
the timeline, let's look at the overarching philosophy
16:27
of this workflow. We covered a massive amount
16:28
of technical ground today. It is a dense, demanding
16:31
process. Claude establishes the mathematical
16:34
visual rules. Chachi PT 5 .6 builds the persistent
16:38
physical assets. And Google Flow brings those
16:41
assets to life with simulated physics. The core
16:44
takeaway for you is this. True realism in AI
16:47
video creation does not come from blindly trusting
16:51
a single magical platform. The hype cycle wants
16:54
you to believe it does, but it never does. Realism
16:58
comes from enforcing rigid stylistic consistency
17:01
across multiple tools. It requires using incredibly
17:05
strong reference images, and it demands taking
17:08
modular step -by -step control of the movement.
17:11
You have to guide the machine. It cannot guide
17:13
you. It gives you so much more control than just
17:15
testing random video platforms and burning credits.
17:18
You stop hoping one magical prompt produces a
17:20
finished masterpiece. You engineer the result
17:23
yourself, brick by brick. You become an actual
17:25
director, not just a prompter. You command the
17:27
lighting, you cast the actors, and you enforce
17:29
the physics. I highly recommend you try this
17:31
exact to stack on your next project. Just to
17:34
see it work, take a photo from your own camera
17:36
roll, run it through Claude Opus 5, just to see
17:39
how accurately the AI dissects the visual style
17:42
of your own life. It will surprise you. It is
17:44
a phenomenal exercise. It really helps you understand
17:46
how the machine actually sees the parameters
17:49
of the physical world. But I want to leave you
17:51
with a final thought today, a philosophical question.
17:53
We are using three separate tools right now,
17:57
but these tools will inevitably merge into a
18:00
single seamless interface. It is just a matter
18:03
of time before Claude, ChatGPP and Flow are one
18:06
button. That convergence is definitely happening
18:09
soon. The friction will disappear. So when that
18:11
happens, will the role of a video creator shift
18:14
entirely? Will we move completely away from crafting
18:17
a scene from scratch to merely curating and correcting
18:20
a simulated reality that the machine builds around
18:23
us? It is a profound shift in human creativity.
18:26
We transition from artists to editors of synthetic
18:29
reality. Something to mull over while you wait
18:31
for your next generation to render. Thanks for
18:34
joining us on this deep dive. Stay curious.