Latest / AI Fire Daily / #45 Robin: The AI Film Masterclass - Why Giant Prompts Fail & The Seedance Continuity Workflow
Transcript
- 0:00Think about the pure chaos of a machine trying
- 0:02to dream. Oh, absolutely. It gets messy fast.
- 0:05You picture a cinematic masterpiece in your head.
- 0:07You type it out with perfect clarity. You hit
- 0:09generate. And what you get back is a total fever
- 0:13dream. Faces morph into strange... unrecognizable
- 0:17shapes. Right. And cars turn into these unidentifiable
- 0:20melting blobs. The action sequence makes absolutely
- 0:23no physical sense. It really does. The continuity
- 0:26is just entirely broken. The lighting shifts.
- 0:29The camera angles just defy gravity. But it honestly
- 0:32doesn't have to be that way anymore. I recently
- 0:34watched a truly seamless AI short film. It had
- 0:38perfect locked -in continuity. Wait, really?
- 0:41Perfect continuity? Yeah. It featured a moving
- 0:44armored convoy with real physical weight. It
- 0:46even had baked -in audio. The dialogue actually
- 0:49matched the lip movements perfectly. It feels
- 0:51like actual magic when it finally works. Welcome
- 0:53to another Deep Tech. Glad to be here. Today,
- 0:56we are exploring something very specific and
- 0:58very powerful. We are unpacking the reference
- 1:01-first AI short film workflow. Our mission is
- 1:05to guide you through this exact process. We want
- 1:09to bridge that gap between chaotic AI generation
- 1:11and true cinematic control. We are going to look
- 1:14at a few critical steps today. We will cover
- 1:17the beginner's mega prompt mistake. We will talk
- 1:20about building solid visual references. We will
- 1:23discuss generating the actual video files. Then
- 1:26we tackle maintaining that incredibly tricky
- 1:29continuity. Finally, we look at the simple editing
- 1:32phase. Let's unpack this core principle first.
- 1:34The idea of control versus chaos. Most beginners
- 1:37start entirely in the wrong place. They jump
- 1:39straight into the video generator. Yeah, and
- 1:41it is a guaranteed recipe for immediate failure.
- 1:44I think we need to understand the psychology
- 1:45here. We want the technology to be magic. So
- 1:48beginners write one giant, overly detailed prompt.
- 1:51Oh, the mega prompt. Exactly. They type something
- 1:53like serious soldier on a desert bridge. They
- 1:55add explosions, dramatic lighting, fast Hollywood
- 1:58pacing. They just expect the AI to handle the
- 2:01rest of the movie. But text alone cannot lock
- 2:04down complex visuals. Text is essentially a low
- 2:07bandwidth communication method. The AI engine
- 2:11is basically just guessing at the details. Because
- 2:13it doesn't actually know what you want. Right.
- 2:15It has to invent the character's facial structure.
- 2:17It guesses the outfit, the specific location,
- 2:20and the background props. It even has to guess
- 2:22the timing and the physics. I have a vulnerable
- 2:25admission to make here. I still wrestle with
- 2:27prompt drift myself. It's so tempting to just
- 2:30ask the AI to read my mind. We all do it. We
- 2:33want that immediate dopamine hit. We want the
- 2:35quick shortcut. But skipping the foundational
- 2:38work always costs you hours of time later. It's
- 2:41kind of like giving a chef a list of ingredients
- 2:43but no recipe. You just expect a Michelin star
- 2:46meal to magically appear. That is a perfect analogy.
- 2:49Let's talk about the mechanics of why that fails.
- 2:51The AI is a probabilistic engine, right? It starts
- 2:54with pure visual noise. It resolves that static
- 2:57into an image based on your text. Yes, step by
- 3:00step. Exactly. If your text just says serious
- 3:03soldier, there are millions of soldiers in its
- 3:05latent space. The AI picks one variation entirely
- 3:08at random. In the very next frame, it picks a
- 3:12slightly different variation. And that leads
- 3:14to the shifting visuals you mentioned. Clothes
- 3:16change color across different cliffs. Cars just
- 3:19mysteriously disappear from the road. The model
- 3:22simply forgets what it just drew. So by front
- 3:24-loading the visuals, we're basically taking
- 3:26all the dangerous guesswork away from the AI
- 3:29engine. Precisely. You lock the visual parameters
- 3:31down tight. The AI no longer has to invent every
- 3:34single pixel from scratch. Right. We define the
- 3:36world first so the AI doesn't have to. Exactly.
- 3:39Two -sec silence. So let's talk about how to
- 3:42actually do that. This brings us to step one
- 3:44of the workflow. You need to meticulously build
- 3:47your foundational assets. We are talking about
- 3:50the holy trinity of film. Characters, locations,
- 3:53and props. Yes, and you want to use tools specifically
- 3:56designed for cinematic results. The source material
- 3:59recommends using Higgs field inside Cinema Studio.
- 4:02Let's pause and clarify that environment. Cinema
- 4:05Studio is essentially a professional generative
- 4:07workspace. It handles the heavy lifting behind
- 4:10the scenes. You set the model to auto inside
- 4:13that workspace. You choose a 9 .16 aspect ratio
- 4:17for vertical video. And you set the output resolution
- 4:20to a crisp 4K. The auto model is a fascinating
- 4:25choice. It removes unnecessary decision -making
- 4:27from the user. You just describe what you want
- 4:30to see. Yeah, the system automatically routes
- 4:32the prompt to the most suitable underlying generator.
- 4:34It streamlines all the technical friction. Let's
- 4:37talk about building the actual characters next.
- 4:39In this short film example, we have two main
- 4:41guys. We have Ryder. who is our tactical commander.
- 4:44For Ryder, the workflow suggests using a reference
- 4:46image. You start with a stock -style photo of
- 4:49a serious commander. You want his realistic face
- 4:52clearly visible. You know the dark hair and the
- 4:55light stubble. You specify modern black tactical
- 4:57gear. Having that anchor image keeps the character
- 5:00incredibly consistent. The AI studies the facial
- 5:03geometry. But what about the other character?
- 5:06We have a snoper named Vance. For Vance, the
- 5:09approach is entirely different. You don't start
- 5:11with a downloaded reference image. You use an
- 5:13integrated tool called AI Cast instead. AI Cast
- 5:17is incredibly powerful for world building. Okay,
- 5:20let's define that term for clarity. What exactly
- 5:22is AI Cast? A preset menu to build character
- 5:25looks without typing long descriptions. That
- 5:28sounds incredibly useful for maintaining sanity.
- 5:31You just pick the genre from a drop -down, like
- 5:33action. Yeah, you even set a virtual production
- 5:35budget, like, say, $200 million. That budget
- 5:38setting tells the AI to aim for high -end aesthetics.
- 5:41You choose an archetype, like the sage. You pick
- 5:44a white male in his 40s. Right, and you can seamlessly
- 5:47add details, like facial scars or a rugged beard.
- 5:50It feels much more like casting a real human
- 5:52actor. Beat, that covers the characters. Now
- 5:55we need to establish the location. The main scene
- 5:57happens on a highly cinematic desert bridge.
- 5:59You use the cinematic locations mode for this
- 6:02specific task. You describe a wide, imposing
- 6:05desert bridge stretching over a deep canyon.
- 6:08You specify harsh noon sunlight. You add a dusty,
- 6:11atmospheric haze to the air. You generate that
- 6:14pristine, clean version of the bridge. But here
- 6:17is where the workflow gets really interesting.
- 6:20You also have to create a mathematically damaged
- 6:22version of it. This step is absolutely crucial
- 6:25for the later action scenes. You upload that
- 6:28clean image right back into the system. You explicitly
- 6:31ask the AI to add severe explosion damage. You
- 6:35want broken structural steel. deep burn marks,
- 6:39and lingering smoke, but you give the engine
- 6:41one non -negotiable instruction. Do not change
- 6:44the original bridge structure. Exactly. Whoa.
- 6:48Imagine generating the exact explosion damage
- 6:50before the explosion even happens. It genuinely
- 6:53feels like you are cheating time. You are building
- 6:55the future aftermath before the event occurs,
- 6:57but it is absolutely essential for the logic
- 7:00of the workflow. Why not just let the video model
- 7:02figure out what a destroyed bridge looks like
- 7:04when the bomb goes off? Well, if you let the
- 7:06video model invent the destruction dynamically,
- 7:08it loses context. It might completely redesign
- 7:12the entire canyon in the background. It might
- 7:14accidentally change the time of day. You lose
- 7:16the continuity. Exactly. Because the AI might
- 7:19generate a completely different bridge altogether.
- 7:21That makes sense. Finally, you need to establish
- 7:24your reusable props. In this specific case, it
- 7:27is a heavy armored convoy car. You want a clean,
- 7:30side -front view of this exact vehicle. You use
- 7:33a highly realistic model for this generation.
- 7:36The source suggests something sophisticated,
- 7:38like Soul Cinema 2K. You just want a clear, well
- 7:42-lit image of the physical object, heavy military
- 7:45design elements, reinforced bulletproof windows,
- 7:49rugged oversized tires. You save that image alongside
- 7:52your other visual references. Let's quickly review
- 7:55the assets we have gathered. We have Ryder, we
- 7:57have Vance, we have a pristine, clean bridge,
- 7:59we have a heavily damaged bridge, and we have
- 8:02our armored car. That structural logic is exactly
- 8:05what we need. Beat. We will get into the video
- 8:08generation phase right after this quick break.
- 8:11Sponsor. All right, so we have our visual bays
- 8:13officially locked. We are finally ready for step
- 8:15two. generating the actual video sequence. This
- 8:18is where we move from static pictures to fluid
- 8:20moving scenes. We use a very smart prompt structure
- 8:24for this phase. You open your dedicated AI movie
- 8:27generator. The source material uses a platform
- 8:29called C -Dance. You start by opening the director
- 8:32panel. You set the overarching genre to action.
- 8:35You choose smart for the integrated shot control.
- 8:39Smart control is infinitely better for beginners.
- 8:41It handles the complex camera angles and virtual
- 8:44movement automatically. Because manual camera
- 8:46control involves deep math. It's just too complex
- 8:50at first. Yeah, exactly. You set the clip duration
- 8:52to 10 or 15 seconds, and you critically turn
- 8:55the audio on in. That audio setting is a massive
- 8:58paradigm shift. Then you feed your visual references
- 9:01into the AI system. You upload the images of
- 9:04Ryder, Vance, the clean bridge, and the armored
- 9:07car. This shift in the workflow is huge. You
- 9:10are no longer desperately describing writer with
- 9:12text. You are actively selecting writer from
- 9:14your pre -built assets. Next, you have to carefully
- 9:17set the character emotions. You click the little
- 9:18parameter icon next to their faces. For the quiet
- 9:21opening scene, you choose medium vigilance. Beginners
- 9:24almost always push the emotional intensity too
- 9:27high. They choose maximum rage or sheer panic
- 9:29right away. Right, but high intensity looks incredibly
- 9:33cartoonish and dramatic for a quiet setup. A
- 9:36medium emotion level feels so much more grounded
- 9:38and natural. Beat. Then we get to the actual
- 9:42text prompt. We use something called the multi
- 9:44-shot framework. You break the written prompt
- 9:46into clear, numbered shots. Shot one, shot two,
- 9:49shot three. You describe the specific camera
- 9:52action for each individual shot. And you put
- 9:54the spoken dialogue lines directly in quotation
- 9:57marks. It is brilliant. Instead of throwing a
- 9:59whole bucket of text at the wall, it's like stacking
- 10:01Lego blocks of data. That is exactly how it feels.
- 10:05It is highly structured and remarkably orderly.
- 10:07Shot one is the establishing wide shot of the
- 10:10bridge. Shot two focuses on Ryder crouching behind
- 10:14tactical cover. Shot three reveals Vance locked
- 10:17in his high sniper position. And then shot four
- 10:20delivers the actual dialogue. Ryder says, everyone
- 10:23in position. The AI follows this logical sequence
- 10:26step by step. It doesn't get overwhelmed and
- 10:28confused by a massive wall of text. If we leave
- 10:30the audio on end during generation, are we getting
- 10:33Actual usable dialogue right out of the box.
- 10:35Yes, the tool generates the vocal dialogue internally.
- 10:38It synthesizes the voice and adds ambient environmental
- 10:41sounds. It bakes those basic effects right into
- 10:45the final video file. So you don't have to record
- 10:47separate voice lines later. Exactly. Yeah, so
- 10:49it bakes in the voices and sound effects automatically.
- 10:51Huge time saver. Two sec silence. We generated
- 10:56our first successful scene, but one cool clip
- 10:59doesn't make a coherent movie. Step three is
- 11:02keeping every subsequent clip logically connected.
- 11:05This is exactly where most AI films completely
- 11:08fall apart. You string two generated clips together
- 11:10and the cut feels entirely wrong. The underlying
- 11:14mood suddenly resets to zero. People usually
- 11:17try the outdated traditional method first. They
- 11:19take the very last frame of clip one. They use
- 11:22that single image to start clip two. But that
- 11:24methodology fundamentally fails. A single still
- 11:27image cannot carry emotional tension. It only
- 11:30captures one static, frozen moment in time. It
- 11:33has absolutely no velocity. You need the next
- 11:36clip to implicitly remember what just physically
- 11:38happened. That is exactly why you use video reference
- 11:40instead. Let's define that critical concept quickly.
- 11:43What is video reference? Feeding the whole previous
- 11:46video to keep the exact same mood. So the AI
- 11:48mathematically reads the scene's pacing. It reads
- 11:51the ongoing camera motion. It reads the lingering
- 11:54tension from the previous scene. Let's dig into
- 11:57the mechanics of that. How does it actually read
- 11:59tension? Well, it reads the pixel movement and
- 12:02the audio waveforms. It analyzes the speed and
- 12:05direction of the motion vectors. So yes, it mathematically
- 12:08calculates and continues the kinetic energy.
- 12:10Clip two uses clip one as its direct video reference.
- 12:13This is our big explosion scene. You change writer's
- 12:17underlying emotion from vigilance to pure rage.
- 12:20You keep Vance firmly on vigilance, and you finally
- 12:23swap in the damaged bridge reference image. You
- 12:25write the next multi -shot prompt sequence. Ryder
- 12:29aggressively triggers the remote explosion. The
- 12:31bridge violently erupts in thick smoke and heavy
- 12:34debris. The scene carries forward incredibly
- 12:36smoothly. Then we get to clip three. This is
- 12:39easily the hardest action scene to generate.
- 12:41It is an absolute chaotic battle zone. Multiple
- 12:45characters are moving simultaneously. Heavy vehicles
- 12:49are reacting to the blast. Debris is falling
- 12:52everywhere across the frame. Because the visual
- 12:55data is so dense, the source material suggests
- 12:57a specific tactic. You bump the total generations
- 13:01up to three. I want to understand this. Why do
- 13:05we need more generations here? You fundamentally
- 13:07need better statistical odds. Dense action has
- 13:10so many overlapping moving parts. The AI doesn't
- 13:13truly understand physics. It just predicts pixels.
- 13:16Right. One take might miss the explosive timing
- 13:18completely. Another might accidentally block
- 13:20the main character's face. So for that messy
- 13:22action scene, we aren't expecting perfection,
- 13:25just asking for a few takes to choose the best
- 13:27one. Precisely. You watch all three generated
- 13:29variations and pick the clearest, most accurate
- 13:31take. Exactly. We're just playing the odds when
- 13:34the action gets heavy. Then you move on to clip
- 13:35four. This is the narrative ending. You need
- 13:38to deliberately slow the visual pacing down.
- 13:41You drop the generations back down to one. The
- 13:43immediate physical threat is almost entirely
- 13:45over. The emotional shift in this phase is very
- 13:49important. You change Vance's baseline emotion
- 13:51from vigilance to trust. If he visibly stays
- 13:55tense, the ending feels completely unresolved.
- 13:58The trust emotion physically tells the AI to
- 14:01relax the character's posture. Thick smoke slowly
- 14:04drifts across the ruined bridge. Ryder cautiously
- 14:07lowers his tactical weapon. Vance finally relaxes
- 14:11his grip on the high cliff. Ryder simply says,
- 14:13bridge is clear. The narrative sequence is completely
- 14:16resolved. Now you have four deeply connected
- 14:18clips. They logically follow a continuous physical
- 14:22path. They actually feel like a real directed
- 14:24sequence. Which naturally brings us to step four.
- 14:27You have to finally edit and export the final
- 14:30film. You assemble the final product in a program
- 14:32like CapCut. It is incredibly simple and fast
- 14:35for beginners to use. You import the generated
- 14:37clips in exact chronological order. Yeah. The
- 14:39quiet setup. The massive explosion. the chaotic
- 14:43action, the calm end. You watch the full timeline
- 14:46sequentially to feel the narrative rhythm, but
- 14:49you keep the actual edits incredibly simple.
- 14:51Wait, let me push back on this. So we are barely
- 14:53editing at all. That goes against everything
- 14:55I know about traditional video production. The
- 14:58edit bay is usually where the movie is actually
- 15:00made. I completely understand that reaction,
- 15:02but the video reference step already did the
- 15:05incredibly hard work. Yeah. That is the real
- 15:07paradigm shift here. Okay. The lighting and the
- 15:10atmospheric mood already match perfectly. The
- 15:13dialogue audio is already seamlessly baked in.
- 15:15Because we controlled everything in the reference
- 15:17and generation phases, the edit bay is just for
- 15:20polishing. Yes, editing simply becomes basic
- 15:24superficial cleanup. You are no longer desperately
- 15:27trying to save a broken, disjointed film. The
- 15:30real heavy lifting happens long before we ever
- 15:33open CapCut. You simply trim the slightly weak
- 15:36clip openings. You cut any awkwardly long pauses.
- 15:39You carefully remove any visually broken frame.
- 15:42You absolutely do not add heavy, distracting
- 15:44visual transitions. Then you hit export. You
- 15:46use a 16 .9 ratio for a classic cinematic look,
- 15:49or you use 9 .16 for modern vertical platforms.
- 15:53You gotta always watch it outside the editor
- 15:54first. on your phone, you will catch small continuity
- 15:57problems or pacing issues much easier that way.
- 15:59Two sec silence. So what does this all essentially
- 16:02mean? Let's recap the big idea here. The main
- 16:04takeaway is entirely about intention and control.
- 16:07Do not ever make the AI guess your movie. You
- 16:11show it exactly what your specific world looks
- 16:13like first. Only then do you tell it what should
- 16:16actually happen. The magic formula is quite clear
- 16:19and replicable. Solid references, structured
- 16:22video generation, mathematical continuity, simple
- 16:25edit. You meticulously build your characters
- 16:28and locations. You write a smart multi -shot
- 16:31prompt sequence. You mathematically connect clips
- 16:33with video references. You easily clean it up
- 16:36in CapCut. It completely changes how you approach
- 16:39these generative AI tools. It turns a folder
- 16:41of random chaotic clips into a real watchable
- 16:45film. It is a profound shift in mindset. You
- 16:48become a deliberate director, not just a passive
- 16:50prompter. Before we sign off, I want to leave
- 16:53you with a final thought to mull over. I always
- 16:55love these conceptual questions. Let's hear it.
- 16:57We just spent this entire time talking about
- 16:59how perfectly an AI can maintain a fictional
- 17:01world. It mathematically does it just from a
- 17:04few carefully constructed reference shots. But
- 17:07what happens when we inevitably start feeding
- 17:09it visual references of our actual lives? Pictures
- 17:13of our real homes, audio of our past memories.
- 17:16How long until the films we casually generate
- 17:18are completely indistinguishable from the lives
- 17:21we've actually lived? That is a fascinating and
- 17:23perhaps slightly terrifying question to consider.
- 17:26The line between generated memory and real memory
- 17:29is getting very thin. Thank you for diving deep
- 17:32with us today. Keep learning, keep experimenting
- 17:34with these tools, and we'll see you on the next
- 17:36deep dive. Out to your own music.