Latest / AI Fire Daily / #93 Robin: Stop Making Plastic AI Videos - How to Actually Direct Seedance 2.5 with Claude Skills

Transcript

#93 Robin: Stop Making Plastic AI Videos - How to Actually Direct Seedance 2.5 with Claude Skills · AI Fire Daily

• 0:00
  You've just generated this gorgeous cinematic
• 0:02
  AI video. It looks completely flawless for exactly
• 0:06
  five seconds. Right. But then your main character
• 0:08
  leans on a washing machine. The camera drifts
• 0:11
  just a fraction of an inch. Suddenly, the machine
• 0:14
  morphs into something totally unrecognizable.
• 0:17
  Yeah, it completely shatters the illusion. It
• 0:19
  really does. You're beat. Welcome to this deep
• 0:22
  dive. We're exploring a truly fascinating guide
• 0:25
  for you today. Our mission is to dissect exactly
• 0:28
  how to build consistent cinematic AI video. We'll
• 0:31
  be using Sedan's 2 .5 and Claude skills. completely
• 0:37
  shift our creative mindset here. We need to move
• 0:39
  away from hoping for some magic prompt. It's
• 0:42
  all about adopting a highly disciplined film
• 0:44
  director mindset. Because before we can direct
• 0:47
  a masterpiece, we need a testing ground. We need
• 0:49
  something that brutally exposes the AI's weaknesses.
• 0:52
  We really have to start slow. That's a crucial
• 0:54
  point to understand right away. The guide introduces
• 0:57
  a very specific demo project for this. It's called
• 0:59
  Control Wash. It isn't some heavy action scene
• 1:02
  with massive explosions. There are absolutely
• 1:05
  no wild car chases happening here. It's just
• 1:08
  a simple, incredibly slow scene inside a laundromat.
• 1:12
  I love how straightforward the plot actually
• 1:14
  is. Picture this. A character named Tay is totally
• 1:19
  exhausted doing his laundry late at night. Right.
• 1:21
  Suddenly, a person in a full gorilla suit wanders
• 1:24
  in. They're casually carrying a large laundry
• 1:27
  basket. Tay looks confused. The notice is something
• 1:30
  very specific. He sees a discarded AI sloppy
• 1:34
  bar chocolate wrapper. It sounds almost too simple,
• 1:37
  you know, like, why use this for technical testing?
• 1:40
  But this kind of scene actually matters deeply.
• 1:43
  Slow scenes expose generative problems incredibly
• 1:45
  fast. Why is exactly why do slow scenes broke
• 1:48
  the illusion so easily? Well, because AI video
• 1:50
  models don't actually understand what a washing
• 1:52
  machine is. They don't have a 3D engine running
• 1:54
  inside their neural network. They're literally
• 1:56
  just guessing based on flat pixel patterns. Wow.
• 1:58
  Yeah, so if Tay moves to a completely different
• 2:00
  washer, you notice. If the camera angle changes
• 2:02
  slightly, it just feels fake. Our brains are
• 2:04
  just highly tuned to spatial consistency. We
• 2:07
  want every single clip to feel perfectly grounded,
• 2:10
  like it was actually filmed in the same physical
• 2:12
  room. Exactly. And many people try using a start
• 2:15
  frame image for every single shot. They think
• 2:17
  it keeps the visual consistency perfectly intact.
• 2:20
  But doing that just makes the final video look
• 2:23
  far too plastic. The actual goal is something
• 2:25
  organic, messy, and real. It's exactly like stacking
• 2:29
  Lego blocks of data. If your foundational base
• 2:31
  play is wobbly or artificial, nothing works.
• 2:34
  The whole visual tower eventually falls apart,
• 2:37
  beat. Do slow scenes actually stress -test the
• 2:41
  AI's spatial memory harder than fast, chaotic
• 2:43
  action? Yes, because the human eye has time to
• 2:46
  linger on the geometry. In a fast action scene,
• 2:49
  motion blur hides a lot of generative mistakes.
• 2:52
  In a slow scene, the model has to continuously
• 2:54
  prove it understands the physical room. Right.
• 2:56
  Less motion means those visual mistakes just
• 2:59
  can't hide. So if we aren't relying on rigid
• 3:02
  start frames to keep the scene stable, what do
• 3:04
  we do? How do we actually build that sturdy Lego
• 3:07
  base plate? We have to program the AI's brain
• 3:09
  before we ask for a video. And this brings us
• 3:13
  directly to the pre -production workflow. You
• 3:15
  don't just ask sedents for a random video prompt.
• 3:19
  You must meticulously build the world first.
• 3:22
  You start by loading specific Claude skills into
• 3:25
  your workflow. And for those newer to this, a
• 3:28
  Claude skills like a custom instruction set.
• 3:30
  It turns a general chatbot into a highly specialized
• 3:33
  planner. Precisely. The guide recommends three
• 3:36
  specific directing tools for this. You load Banana
• 3:39
  Pro Director, Cinema Director V3, and Character
• 3:42
  Builder. These skills give Claude a highly structured,
• 3:45
  rigid directing system. So it stops it from just
• 3:48
  generating pretty meaningless text. Instead,
• 3:51
  we use this to create a scene plate. Can you
• 3:53
  define what a scene plate is for us? The foundational
• 3:56
  background image before characters are added.
• 3:58
  Perfect. So we build the empty location before
• 4:00
  anyone shows up. Let's look closely at the control
• 4:02
  wash scene plate. So it uses a three -quarter
• 4:05
  camera angle for a very specific reason. This
• 4:07
  angle makes the room feel easy to reuse across
• 4:09
  different shots. It gives you depth without locking
• 4:12
  you into a flat wall. Right, and it features
• 4:14
  running washers, which adds subtle background
• 4:16
  life. There's a mirrored neon sign for strong
• 4:19
  visual identity. Plus, the night lighting creates
• 4:22
  a really moody cinematic atmosphere. The tool
• 4:25
  choice here is also very deliberate. You should
• 4:28
  use Cdream 5 .0 Pro for these scene plates. Why
• 4:32
  that specific tool instead of another generator?
• 4:34
  because it adds beautiful, organic, visual imperfections
• 4:39
  to the image. You get slightly overexposed lights
• 4:41
  and a film -like grain glow. It makes the digital
• 4:44
  image feel like an actual camera lens captured
• 4:47
  it. But for the character faces, you switch tools
• 4:49
  entirely. You use Nano Banana Pro specifically
• 4:52
  for the human faces, because human faces require
• 4:54
  much stricter visual consistency than background
• 4:57
  neon signs. Right, because our brains are hardwired
• 4:59
  to notice facial inconsistencies immediately.
• 5:02
  Better input always gives you much better output.
• 5:04
  And this leads to a very specific disciplined
• 5:07
  prompting strategy. You build your character
• 5:09
  consistency sheets using a tool like Higgs Field.
• 5:12
  So Higgs Field acts sort of like a digital casting
• 5:14
  agency for your AI. Yeah, exactly. You get consistent
• 5:18
  character turnarounds to use as references. You
• 5:21
  screenshot those elements and feed them directly
• 5:23
  to Claude. And this is where the prop structure
• 5:25
  becomes absolutely vital. You put the most crucial
• 5:28
  information at the very top. Like what kind of
• 5:30
  information? You explicitly list the shot count
• 5:33
  and total duration right away. You even explicitly
• 5:36
  state no music whatsoever at the beginning. Why
• 5:39
  does the gorilla costume matter so much in the
• 5:42
  prompt? The guide specifically calls out phrasing
• 5:45
  it as a person wearing a gorilla costume. Right.
• 5:49
  And that stops it from animating a real animal.
• 5:51
  If you just say gorilla, the AI animates a quadruped
• 5:55
  walking like an ape. If you say person wearing
• 5:57
  a costume, it forces the AI to animate a human
• 6:01
  walking. Oh, wow. Yeah, it completely changes
• 6:03
  the physical gait of the character. You also
• 6:05
  always start the prompt mid -action, right? So
• 6:07
  the washer should already be moving when the
• 6:09
  video frame begins. Two -sec silence. I'll be
• 6:13
  honest with you, I still wrestle with prompt
• 6:14
  drift myself whenever I get impatient. I always
• 6:17
  try to skip this world -building phase. Incredibly
• 6:19
  common, yeah. It's just so tempting to rush it
• 6:21
  and hit generate. Does putting instructions like
• 6:24
  shot count at the very top of the prompt actually
• 6:26
  change the technical output? Absolutely. Video
• 6:29
  models weigh the first few tokens much more heavily
• 6:31
  than the end of a paragraph. It's a fundamental
• 6:34
  quirk of their architectural attention mechanism.
• 6:37
  Makes sense. The AI prioritizes the very first
• 6:40
  instructions it reads. Beat. Now we have the
• 6:43
  perfect mid -action first five seconds. But how
• 6:46
  do we keep the scene from completely hallucinating
• 6:49
  in second six? This is where video consistency
• 6:51
  becomes the primary daily challenge. Seedens
• 6:54
  2 .5 has a fascinating feature specifically designed
• 6:58
  for this. It's called the Extend Video feature,
• 7:01
  and it's a massive game changer. You can upload
• 7:03
  your generated clips up to 30 seconds long, right?
• 7:05
  Yes, and the model actively reads the character
• 7:08
  placement and visual appearance. It analyzes
• 7:10
  the camera angle and the physical room layout.
• 7:13
  Then, it attempts to generate what logically
• 7:16
  happens next. There are two distinct powerful
• 7:19
  modes for this specific workflow. The sequel
• 7:22
  mode creates what happens logically after the
• 7:24
  clip. And the prequel mode creates what theoretically
• 7:27
  happened before the clip. This allows Tay to
• 7:29
  stay leaning casually on the same washer. He
• 7:32
  holds the exact same physical position from the
• 7:35
  previous shot. The laundromat layout remains
• 7:37
  perfectly intact all around him. But you definitely
• 7:40
  don't just hit a button and hope for the best.
• 7:43
  There's a meticulous, almost grueling, frame
• 7:47
  -by -frame workflow involved here. You drag the
• 7:50
  current video clip into Claude first. You ask
• 7:53
  Claude to analyze the 2D scene deeply. You meticulously
• 7:56
  track character position and specific hand placement.
• 7:59
  You check the camera angle and the distant background
• 8:02
  layout. Exactly. Then you ask Claude to write
• 8:04
  a highly specific continuation prompt. After
• 8:08
  that, you paste that new prompt back into Higgs
• 8:10
  field. The crucial takeaway here is always about
• 8:13
  absolute control. You must explicitly tell the
• 8:16
  AI what must stay locked. You actively lock the
• 8:18
  eye direction and specific hand position. Yeah.
• 8:21
  You lock the background objects and the global
• 8:23
  room lighting. You don't just lazily describe
• 8:25
  what happens next in the story. You have to tell
• 8:27
  it what doesn't change. Yes, that's the secret.
• 8:29
  Two -sec silence. Whoa! Imagine the compute power
• 8:33
  required to have an AI analyze its own previous
• 8:35
  video frame by frame just to keep a character's
• 8:39
  hand perfectly locked on a washing machine. That
• 8:41
  is wild. It's a staggering amount of processing
• 8:44
  for sure. Can the model truly understand 3D space
• 8:48
  from these 2D frames or is it just guessing?
• 8:50
  It doesn't possess true 3D spatial awareness.
• 8:54
  It's heavily approximating depth based on pixel
• 8:57
  patterns and object relationships over time.
• 8:59
  So it basically fakes depth by tracking object
• 9:02
  relationships over time. Tracking those tiny
• 9:04
  pixel relationships over time requires incredible
• 9:06
  focus. It takes a massive amount of sustained
• 9:09
  processing power. And when you're deep into learning
• 9:11
  complex systems yourself, you need that same
• 9:13
  level of focus. That's why I rely daily on ClearMind
• 9:16
  Coffee. Oh, nice. It's a small batch roast designed
• 9:19
  for sustained jitter -free concentration. It
• 9:22
  genuinely helps me stay locked in when analyzing
• 9:25
  dense material. Check them out when you need
• 9:27
  that quiet, steady focus... bead. Now, let's
• 9:31
  get back to the deep dive. We really need to
• 9:33
  give you a reality check on resources now. AI
• 9:36
  video generation is incredibly fun and creatively
• 9:39
  freeing, but it can get very expensive very fast
• 9:42
  if you aren't careful. The Simple Control Wash
• 9:44
  project is a perfect example of this. Producing
• 9:46
  that short scene cost around 800 credits on Cdance
• 9:49
  2 .5. 800 credits just for a guy in a laundromat.
• 9:53
  Yeah. That's exactly why writing better AI video
• 9:56
  prompts matters so much. The guide heavily advises
• 9:59
  using unlimited generation promos when starting
• 10:01
  out. They're definitely slower, taking one generation
• 10:04
  at a time, but it saves your valuable credits
• 10:06
  while you actually learn the system. The ultimate
• 10:08
  goal of this meticulous workflow is pure efficiency.
• 10:11
  You want to fundamentally reduce your wasted,
• 10:13
  expensive credits. Right. You rigorously plan
• 10:16
  the shot first, and only then do you generate.
• 10:19
  And this exact same workflow applies to highly
• 10:22
  active shots, too. Think about extreme sand snowboarding
• 10:25
  or a massive dragon lifting off. But active shots
• 10:29
  require incredibly strict physics direction from
• 10:31
  you. Because the AI doesn't inherently understand
• 10:34
  how gravity actually works. Exactly. You have
• 10:37
  to explicitly define the physical weight and
• 10:39
  gravity in the prompt. You must describe how
• 10:41
  the dust settles and how the camera tracks it.
• 10:44
  You have to control the implied speed and the
• 10:46
  character's body balance. You also have to fundamentally
• 10:48
  reframe your view of failures. A failed generation
• 10:51
  isn't entirely useless to your project. Maybe
• 10:54
  a video clip has an absolutely terrible glitchy
• 10:57
  ending, but it might have a truly great flawless
• 10:59
  first three seconds. Yeah, you intentionally
• 11:01
  save those perfect first three seconds. Real
• 11:04
  filming on a physical set requires multiple takes
• 11:06
  and careful editing. You build the final cinematic
• 11:09
  scene slowly from the best captured moments.
• 11:11
  It's exactly like stacking leg... Well, actually,
• 11:14
  it's more like panning for gold. You don't keep
• 11:16
  the whole river. You just edit together the tiny
• 11:17
  nuggets. Yeah. Does having access to an unlimited
• 11:21
  generation promo actually encourage lazy spray
• 11:24
  -and -pray prompting? Yes. People tend to stop
• 11:27
  planning their shots and just hit generate, hoping
• 11:30
  for a lucky accident, which ultimately wastes
• 11:33
  time. Yeah. Infinite tries can actually kill
• 11:35
  your creative discipline, which brings us smoothly
• 11:38
  to the ultimate philosophy of this entire deep
• 11:41
  dive. We really need to synthesize this entire
• 11:44
  workflow for you. Cinematic AI video requires
• 11:47
  fully abandoning a major creative myth. You must
• 11:51
  abandon the idea of prompts as magic words. You
• 11:53
  aren't just generating random beautiful footage
• 11:55
  anymore. No. You're actively building a sturdy,
• 11:59
  reusable scene plate first. You're rigorously
• 12:01
  enforcing strict physical rules on a blind model.
• 12:05
  You're actively managing credits and meticulously
• 12:07
  salvaging the best takes. The AI can successfully
• 12:10
  build the technical pixel -perfect shot, but
• 12:13
  your human brain still has to direct the actual
• 12:15
  movie. You have to bring your own taste and editorial
• 12:18
  choices to the timeline. If we're already using
• 12:20
  AI to perfectly lock spatial geometry, if we
• 12:23
  can simulate a 30 -second continuous photorealistic
• 12:26
  shot of a laundromat based solely on text what
• 12:29
  happens to our perception of historical video
• 12:31
  archives a decade from now, when any given moment
• 12:34
  can be perfectly synthesized with organic film
• 12:36
  grain and imperfect lighting, how will we ever
• 12:38
  definitively prove what was actually filmed and
• 12:41
  what was just brilliantly directed by someone
• 12:43
  in their bedroom? Beat? Keep Question Everything,
• 12:46
  keep exploring the tools, and always stay curious.