Latest / AI Fire Daily / #682 Neil: AI Animation From One Idea With ChatGPT And Google Flow

Transcript

#682 Neil: AI Animation From One Idea With ChatGPT And Google Flow · AI Fire Daily

• 0:00
  Imagine making a 3D motion graphics video that
• 0:03
  looks like it took, you know, hours and hours
• 0:05
  in After Effects using just a few AI prompts.
• 0:09
  Sounds crazy, right? It sounds completely absurd,
• 0:12
  honestly. But it's true. Like, if you told me
• 0:14
  this two years ago, I'd say you were out of your
• 0:15
  mind. Yeah. But it's real. So welcome to today's
• 0:20
  deep dive. We are exploring a pretty fascinating
• 0:23
  blueprint today. Very fascinating. We're looking
• 0:25
  at how to build these stunning, incredibly complex
• 0:28
  3D animations. basically from scratch. Right.
• 0:31
  And you are only using ChatGPT and Google Flow.
• 0:35
  I mean, zero animation experience required, which
• 0:37
  is the crazy part. That is the absolute best
• 0:40
  part, really. You don't need to understand complex
• 0:43
  geometry or anything. You just need a solid idea.
• 0:46
  Exactly. So here is our roadmap for this deep
• 0:49
  dive. First, we will break down how to reverse
• 0:53
  engineer a reference video. Yeah, kind of cloning
• 0:56
  an existing style just to learn the ropes. And
• 1:00
  second, we will conjure a completely original
• 1:02
  3D animation, going from just a raw, blank canvas
• 1:05
  idea to a finished piece. So before we start
• 1:08
  directing our digital movie, we kind of need
• 1:10
  to assemble our crew. Right, because you can't
• 1:11
  shoot a film alone. It's like assembling a digital
• 1:14
  film crew where AI plays every role. That's a
• 1:16
  perfect analogy, yeah. So Chat GPT acts as the
• 1:19
  brain. It's your head writer. It analyzes the
• 1:23
  reference videos, right? Exactly. It analyzes
• 1:25
  videos, writes the storyboards, and, you know,
• 1:28
  crafts the exact hyper -specific prompts you
• 1:30
  need. It's doing all the heavy logistical thinking.
• 1:33
  Yeah. Then you have Google Flow. So Flow acts
• 1:36
  as the muscle. It's your actual animator. It
• 1:39
  generates the motion graphics from the prompts.
• 1:41
  Right. And it uses reference images to stay on
• 1:43
  track. And the best part is, you know, you can
• 1:46
  start with free credits. That makes the barrier
• 1:48
  to entry so low, but we also need, like, an editing
• 1:50
  suite. Yeah, just something straightforward,
• 1:52
  CapCut or Premiere Pro. That's your editing bay
• 1:55
  to connect clips. And to sync audio. Exactly.
• 1:58
  And there is also 11 Labs, but... That's optional,
• 2:01
  right? Yeah, that's only used if you want to
• 2:02
  add a synthetic voiceover. We're mostly focused
• 2:05
  on the visuals today. If Google Flow actually
• 2:08
  makes the video, why do we even need chat GPT?
• 2:11
  Ah, well, because chat GPT structures the chaos.
• 2:15
  Video models, they operate in this massive latent
• 2:17
  space. They'll just guess blindly. Exactly. Flow
• 2:20
  needs extremely precise, almost mathematical
• 2:23
  directions to look good. Otherwise, it just hallucinates.
• 2:27
  Ah, so chat GPT plans the exact shots before
• 2:30
  Flow animates them. Precisely. You need the architect
• 2:32
  before you hire the builder. So now that we have
• 2:36
  our crew, let's give them a script. The easiest
• 2:39
  way to start is by mimicking something that already
• 2:41
  works. Right. We're stealing like an artist.
• 2:44
  So you start by finding a short 3D motion graphic
• 2:46
  online. On Instagram or TikTok or Pinterest?
• 2:50
  Yeah, you find something with a cool camera move,
• 2:52
  but you want it short, like 10 to 20 seconds
• 2:54
  max. Why is that? Just to save flow credits?
• 2:57
  That's party, but yeah. Shorter saves credits,
• 2:59
  but it's also way easier for the AI to analyze.
• 3:02
  Too many variables in a long video. Exactly.
• 3:05
  So you feed this short video to ChatGPT using
• 3:07
  a really strict three -phase prompt. This is
• 3:11
  crucial. Phase one is the analysis phase. Right.
• 3:15
  You ask it to identify the duration, the aspect
• 3:18
  ratio, the camera moves. It also needs to define
• 3:20
  the color palette, right? Yes. Color palette
• 3:22
  transitions and pacing. If you don't lock down
• 3:25
  the colors early, the AI will just, you know,
• 3:28
  invent things. Then we move to phase two, which
• 3:30
  is the breakdown. Yeah. Here, you tell ChatGPT
• 3:33
  to create a table. It maps out scene numbers,
• 3:36
  exact timestamps, and visual descriptions. It
• 3:39
  also lists out required reference assets. Mm
• 3:42
  -hmm. And finally, phase three. the production
• 3:45
  strategy. ChatGPT kind of becomes a line producer
• 3:48
  here. Exactly. It recommends the number of Google
• 3:51
  flow clips you're going to need and their durations.
• 3:54
  And it flags visual consistency problems. Let's
• 3:57
  look at a specific example from the source material.
• 3:59
  OK, yeah. There's this reference video of a 3D
• 4:03
  robot. Right. And ChatGPT perfectly spots the
• 4:06
  glowing blue eyes. And the white metallic armor.
• 4:09
  It notes all of that across the different shots.
• 4:11
  I still wrestle with prompt drift myself where
• 4:13
  a character morphs unexpectedly, so getting this
• 4:16
  exact breakdown is huge. It's massive. It keeps
• 4:19
  the AI completely focused so the robot doesn't
• 4:22
  suddenly turn into a toaster. How does the AI
• 4:26
  handle changing camera angles from a reference?
• 4:28
  Well, ChatGPT notes exactly when the camera pushes
• 4:31
  in or pans so the prompt reflects it. Got it.
• 4:35
  ChatGPT maps camera moves so flow can match them
• 4:38
  exactly. Right. So we have the architectural
• 4:42
  blueprint from our analysis. Now it's time to
• 4:45
  lay the bricks. We prompt ChatGPT to act as a
• 4:48
  Motion Graphics Director now to build a storyboard
• 4:51
  and exact Google Flow prompts. Yes, but there
• 4:54
  is a crucial rule here. The prompt must absolutely
• 4:57
  separate production instructions from on -screen
• 5:00
  text. Because otherwise the technical notes might
• 5:02
  accidentally appear in the final video, right?
• 5:04
  Exactly. If the prompt says, camera pans left,
• 5:08
  the AI might literally spell those words on the
• 5:10
  screen. Which ruins the illusion. So let's look
• 5:14
  at a 20 -second animation broken into three clips.
• 5:17
  OK, so clip S1. This is a robot close -up with
• 5:20
  a slow camera push -in. Then clip S2 is the robot
• 5:24
  flying through the sky using foot thrusters.
• 5:26
  Right. In clip S3, the camera is moving closer
• 5:30
  to the main robot standing in a group. So we
• 5:32
  take these specific prompts to Google Flow. We
• 5:35
  set the generation settings. We use a 9 .16 aspect
• 5:38
  ratio for a vertical video. And 720p just for
• 5:42
  testing. We set it to 8 seconds per generation.
• 5:45
  Right. Yes. And here's the key. You must provide
• 5:48
  the reference image of the robot to Google Flow
• 5:50
  for every single clip. That anchors the character
• 5:53
  design. Exactly. Then you take it all into CapCut
• 5:56
  for assembly. You trim S1 and S3 down to 6 seconds.
• 6:00
  Yeah. Cut the messy edges off. Leave S2 at the
• 6:03
  full 8 seconds and connect them with straight
• 6:05
  cuts. It's like stacking Lego blocks of data.
• 6:09
  Each 8 -second clip snaps together to form the
• 6:11
  narrative. That's exactly how it feels. You're
• 6:14
  just snapping rendered blocks together on a timeline.
• 6:16
  What happens if one of the robot's thrusters
• 6:18
  looks glitchy in just one scene? Oh, you only
• 6:21
  redo that specific 8 -second chunk, not the whole
• 6:23
  thing. You just regenerate the single bad clip
• 6:25
  instead of the whole video? Yeah, it saves so
• 6:28
  much time and computing power. Sponsor. This
• 6:31
  deep dive is brought to you by our sponsors.
• 6:34
  So... Cloning a reference video is a great training
• 6:36
  ground. It really is. It teaches you how the
• 6:39
  machine thinks. But what if you want to build
• 6:41
  an entirely original world from scratch? That's
• 6:44
  method two. This is where you go from imagination
• 6:47
  to a locked script. You start with a simple idea.
• 6:50
  In the source concept, it's a futuristic smartphone.
• 6:53
  Yeah, it's floating in a dark studio. And then
• 6:56
  mechanical parts start unfolding. It physically
• 6:59
  transforms into a humanoid robot. Well, imagine
• 7:01
  watching a simple phone unfold into a robot right
• 7:04
  before your eyes just from text. It is wild.
• 7:07
  So we prompt ChatGPT as a cinematic 3D animation
• 7:11
  director. We give it a target of 20 seconds,
• 7:14
  a 16 by 9 ratio, a dark futuristic studio. Right.
• 7:18
  We want metallic silver, black, and electric
• 7:21
  blue colors, and no dialogue. Right. So ChatGPT
• 7:25
  takes that and creates a precise six scene script.
• 7:29
  It maps out the exact mechanical transformation.
• 7:32
  And it intentionally carries the smartphone's
• 7:34
  metallic finish and the camera details into the
• 7:37
  robot's final design. So the transformation actually
• 7:39
  makes logical sense. Exactly. The materials don't
• 7:42
  just vanish into thin air. How does the AI know
• 7:45
  how a phone should physically unfold into arms
• 7:48
  and legs? Well, the AI leans on the detailed
• 7:51
  prompts requesting realistic mechanical joints
• 7:54
  and consistent materials. So the AI relies on
• 7:57
  our prompt specifying realistic mechanical components.
• 8:00
  Right. And that specificity grounds the physics,
• 8:02
  so it doesn't just turn into a blurry mess. The
• 8:05
  script is locked. Now we need to render this
• 8:07
  new reality without it turning into a glitchy
• 8:09
  mess. So we ask ChatGPT for character design
• 8:12
  specs. Dimensions, joints, lighting. And we also
• 8:15
  ask it to generate an eight -panel visual storyboard.
• 8:18
  Yes. And we're using that 16 .9 style. Which,
• 8:21
  just to clarify for the listener, aspect ratio,
• 8:24
  the proportional relationship between videos
• 8:27
  width and height. Perfect. So we go back to Google
• 8:30
  flow, 16 .9 ratio, 720p for testing. And we upload
• 8:36
  that eight panel storyboard as the visual reference.
• 8:39
  Right. Then we paste the master prompt that combines
• 8:42
  all the scenes. But what if complex transformations
• 8:45
  break? Like mechanical parts just disappearing.
• 8:49
  If that happens, you generate those specific
• 8:51
  scenes separately, using matching starting and
• 8:54
  ending frames. Ugh. To force a smooth visual
• 8:57
  transition. And then combine everything in CapCut.
• 9:00
  Exactly. CapCut locks in the final timing. Why
• 9:03
  do we upload the 8 -panel storyboard into Google
• 9:05
  Flow instead of just using text? Because text
• 9:08
  alone leaves too much room for interpretation.
• 9:10
  The image really grounds it. It acts as a strict
• 9:13
  visual guide to stop the AI hallucinating. Yeah.
• 9:15
  It's incredible when you step back and look at
• 9:17
  it. We've gone from cloning a clip to directing
• 9:20
  an original VFX piece. What does this workflow
• 9:23
  actually represent for us as creators? It means
• 9:25
  the barrier to entry for high -end motion graphics
• 9:28
  is basically gone. Completely vaporized. ChatGPT
• 9:32
  handles the complex logistical planning. And
• 9:35
  Google Flow acts as your personal render farm.
• 9:37
  You are no longer an animator manually pushing
• 9:39
  keyframes. You are the creative director. You're
• 9:42
  managing a digital crew. You start simple. You
• 9:45
  focus on the movements. And then you scale up.
• 9:47
  The bottleneck is just your imagination now.
• 9:50
  Exactly. We'd encourage you to try this out today.
• 9:53
  Pick a favorite short clip online, run it through
• 9:56
  ChatGPT for a breakdown. Yeah, and just try generating
• 9:59
  one single scene in Google Flow. See how it feels.
• 10:03
  If a single person sitting on their couch can
• 10:05
  conjure up blockbuster level 3D animations in
• 10:08
  a single afternoon, what does the future of visual
• 10:11
  storytelling look like in five years? Will the
• 10:13
  term animation studio just mean one creative
• 10:15
  person with a really good internet connection?