Latest / AI Fire Daily / #490 Neil: AI Video Creation Simplified With 3 Tools And Clear Steps
Transcript
- 0:00You've got this great idea for an AI video. You
- 0:02type it in. You hit generate. And, well, it's
- 0:06a mess. Yeah, a total mess. Your main character's
- 0:09face just keeps morphing. It turns into someone
- 0:11else entirely in the background. Right. The visual
- 0:14style feels incredibly chaotic. What was supposed
- 0:17to be cinematic looks like a complete and utter
- 0:20accident. We have all been there. It's incredibly
- 0:22frustrating. You kind of feel like you're playing
- 0:24a slot machine. You just pull the lever and hope.
- 0:27Welcome to the deep dive. Today we are breaking
- 0:31down a step -by -step guide. We're exploring
- 0:34exactly how to build AI video from scratch. We're
- 0:38looking at a specific three tool sequence. We're
- 0:40talking Cloud AI, ChatGPT, and Google Flow. That's
- 0:44the stack. This sequence guarantees consistent,
- 0:47highly cinematic results. It's really about moving
- 0:49from total randomness to actual directorial control.
- 0:53Right. It really is a mindset shift. We need
- 0:55to understand why those initial generations turn
- 0:57into chaotic messes. It's honestly not about
- 1:01writing a better, more complex prompt. It is
- 1:04entirely about the sequence of events. The sequence.
- 1:06Let's unpack that. The fundamental rule is this.
- 1:09You must build your visuals as static images
- 1:11first. You have to do before animating anything
- 1:14at all. Generating static images is much faster.
- 1:18It's also significantly cheaper than generating
- 1:20video. You know, I still wrestle with prompt
- 1:23drift myself. Oh, yeah. Watching my character
- 1:25morph into a totally different person halfway
- 1:28through a scene, it is a really helpless feeling.
- 1:31Absolutely. Everyone experiences that prompt
- 1:33drift initially. It's like stacking Lego blocks
- 1:35of data. You need a solid, unmoving base before
- 1:38you add the moving pieces. Exactly. If you build
- 1:40on a shaky foundation, the whole scene just collapses.
- 1:42Yeah. But why is that? Why is fixing a mistake
- 1:46in a video so much harder than in a static image?
- 1:50Because video multiplies variables across time.
- 1:53Static images isolate those variables. Think
- 1:55about what a video actually is. Right. A standard
- 1:58film runs at 24 frames per second. If you generate
- 2:01a 10 -second clip, the AI isn't just making one
- 2:04picture. It's making 240 pictures. Exactly. 240
- 2:08distinct images. And here is the truly difficult
- 2:11part for the AI. It has to remember what happened
- 2:14in frame one. It has to apply that exact memory
- 2:18to frame two. Then it has to guess what naturally
- 2:21happens in frame three. That is an immense computational
- 2:24burden. The system has to maintain the physics
- 2:27of a moving scene. It is massive. Eventually
- 2:30the machine drops the ball. The math just gets
- 2:32too heavy. It forgets the color of the character's
- 2:35jacket. It forgets the exact lighting angle on
- 2:37the coffee cup. Wow. But an image generator only
- 2:39has to solve one single frame. It concentrates
- 2:43all its processing power on that exact moment.
- 2:45You get a highly detailed, perfectly accurate
- 2:47picture. So video multiplies the chaos across
- 2:50time. Images isolate the problem entirely. Exactly.
- 2:54It pins the butterfly to the board, so to speak.
- 2:56You lock in the exact visual style. You fix the
- 2:59character's faces. Right. You perfect the lighting
- 3:01before a single frame actually moves. Which means
- 3:03we need a rock solid textual blueprint to start.
- 3:06Yeah. And that brings us to the first tool in
- 3:08this sequence. We're talking about Claude AI.
- 3:11Right. Before you even look at an image generator,
- 3:14before you open a video tool, you need a very
- 3:16clear, structured plan. Claude AI handles this
- 3:20entire blueprint stage. The guide we are looking
- 3:23at uses a very specific example. It starts with
- 3:25a remarkably simple story. A father and daughter
- 3:28escape a sudden volcano eruption. They're driving
- 3:32around desperately searching for gas to survive.
- 3:35It's a great example. It has inherent tension,
- 3:37but visually it's very contained. But let me
- 3:40push back on that a bit. Doesn't giving the AI
- 3:42a highly complex story make for a richer video?
- 3:47Like, if I want a sprawling sci -fi epic, shouldn't
- 3:49my premise be sprawling, too? It sounds counterintuitive,
- 3:52but no. A simple premise is actually much better
- 3:55here. Think about how the AI processes text tokens.
- 3:58OK. If you give the AI a complex novel with 10
- 4:01subplots, it won't know what visual elements
- 4:04to prioritize. Its attention gets scattered.
- 4:06It loses focus on the actual scene. Precisely.
- 4:08Two characters and one clear goal is perfect.
- 4:11It allows the AI to extract very clear visual
- 4:14directions. It knows exactly who to light and
- 4:16what they're doing. Keep the core idea incredibly
- 4:19simple so the AI can extract clear visual directions.
- 4:24Beat. That's the secret. So you take that simple
- 4:27story to Claude, then you create an AI video
- 4:30prompt skill. Let's define that. For someone
- 4:32new to this, what is an AI video prompt skill?
- 4:35Simply put, A saved instruction file for repeated
- 4:39AI tasks. So you aren't just typing into a blank
- 4:41chat box every time? Exactly. You are setting
- 4:43up a permanent framework. You tell Claude how
- 4:46to behave as a master video director. Nice. You
- 4:49load this skill automatically whenever you want
- 4:51to make a video. It saves you from constantly
- 4:53explaining the rules to the AI. What are we actually
- 4:57instructing Claude to do with this skill? What
- 4:59are the outputs we need? This skill demands three
- 5:01very distinct things from your simple story.
- 5:04It forces Claude to break the story down systematically.
- 5:07First, it generates a design sheet prompt. A
- 5:10design sheet? Kind of like concept art for a
- 5:12movie. Right. Exactly like concept art. This
- 5:14prompt outlines the characters, the core environment,
- 5:17and the color palette. It establishes the visual
- 5:20roles. Second, Claude generates a storyboard
- 5:23prompt. This breaks the story down into specific
- 5:27camera angles for each shot. Close -ups, wide
- 5:30shots, panning descriptions. Yes, it acts as
- 5:32the cinematographer. Third, it generates individual
- 5:35video prompts for every single scene. These are
- 5:38highly detailed instructions for the final animation
- 5:41phase. So Claude gives us this perfect structured
- 5:44text. We have the design sheet prompt, the storyboard
- 5:47prompt, and the video prompt. Yeah. But text
- 5:50isn't a movie. We need to actually visualize
- 5:52these blueprints. This is where we transition
- 5:54to the second stage. This is where ChatGPT comes
- 5:56in to lock in the visual style. We're moving
- 5:59from planning to actual image generation. You
- 6:02open ChatGPT. Specifically, you want to use the
- 6:05Deli 3 image generation feature inside it. Your
- 6:08first move is creating that master design sheet.
- 6:10We paste the design sheet prompt that Claude
- 6:12just wrote for us. You do. But the guide emphasizes
- 6:15a crucial addition here. You must manually add
- 6:18a specific phrase to the end. OK. You type. High
- 6:21detail, wide format, every element clearly visible.
- 6:24Why those exact words? It sounds like an arbitrary
- 6:27magic spell. Why not just say, make it look good?
- 6:30Because diffusion models can be lazy. If you
- 6:32just say, make it look good. The AI focuses on
- 6:36aesthetics. It might use heavy shadows to look
- 6:38cinematic. It might blur the background for an
- 6:41artistic depth of field. Which hides the actual
- 6:43details we need. Precisely. We are building a
- 6:46reference document, not a final piece of art.
- 6:50Those specific words force the AI to prioritize
- 6:53spatial clarity. Makes sense. We need flat, even
- 6:56lighting. We need to see the character's exact
- 6:58face. We need to clearly see the buckle on their
- 7:01backpack. This image is the anchor for everything
- 7:03else. So you get this design sheet back, you
- 7:06look it over, but what happens if chat GPT gets
- 7:08a small detail wrong? Say the daughter's backpack
- 7:11is the wrong color, or a photograph on the table
- 7:14is facing the wrong way. The instinct is to just
- 7:18rewrite the whole prompt and try again. That
- 7:20is the biggest mistake people make. Do not rewrite
- 7:22the whole prompt. Why not? Because rewriting
- 7:24changes the underlying seed noise, the AI will
- 7:27generate a completely different image from scratch.
- 7:29Oh, wow. The lighting will change. The faces
- 7:32will change. You lose all the good stuff just
- 7:34to fix one tiny detail. So how do you fix it
- 7:37without destroying the image? You add one specific
- 7:41localized correction sentence. If the photograph
- 7:44is wrong, you literally just type only the back
- 7:47of the photograph is visible. You just replied
- 7:49the image with that one sentence. Yes. The AI
- 7:52understands localized constraints much better.
- 7:55It keeps the original context window intact.
- 7:57It just surgically alters that one specific element
- 8:00you mentioned. Let me play devil's advocate here.
- 8:03Why not just skip this design sheet entirely?
- 8:06Okay. If we have the text prompts, why not generate
- 8:09the storyboard right away. Because without that
- 8:11visual anchor, the AI is just guessing. It will
- 8:14hallucinate a new visual style for every single
- 8:17panel. The dreaded prompt drift again. Exactly.
- 8:20In panel one, your character is wearing a clean
- 8:21blue jacket. In panel two, it's suddenly a dirty
- 8:24denim vest. Right. The lighting shifts from overcast
- 8:27to bright sunlight. It looks amateurish. The
- 8:30design sheet is the exact visual anchor keeping
- 8:33the AI from hallucinating. It absolutely is.
- 8:36So once that design sheet is perfectly locked
- 8:38in, once the character is looking Exactly right.
- 8:40Then, and only then, do you generate the storyboard.
- 8:43You take the storyboard prompt Claude wrote earlier,
- 8:46you paste it into ChatGPT. But here is the critical
- 8:48step. You must attach that perfect design sheet
- 8:52image to the prompt. You are literally feeding
- 8:55the image back into the AI. You tell ChatGPT
- 8:59to match those exact characters and colors. You're
- 9:02giving it visual reference, not just text. Yes.
- 9:04And you add one more specific phrase here. You
- 9:07type, high detail, each panel clearly separated
- 9:10and readable. Because a storyboard is a grid
- 9:13of multiple images. Right. It's usually a two
- 9:15by six grid, 12 panels total. If the AI blends
- 9:18the borders of those panels together, it becomes
- 9:20a mess. You need clean, distinct shots because
- 9:23you will be isolating them soon. We're going
- 9:25to take a quick pause right here. We don't go
- 9:27anywhere. And we are back. So far, we have planned
- 9:31our text in Claude. We have locked in our visual
- 9:34style with a design sheet in ChatGPT. And we
- 9:37just generated a perfectly consistent 12 -panel
- 9:41storyboard. Everything matches. You have a comic
- 9:44book version of your film. It is static, but
- 9:46it is visually perfect. Now comes the magic.
- 9:49We finally breathe life into it. We are moving
- 9:52to the third tool. Google Flow. This is the animation
- 9:56phase. Google Flow uses an underlying technology
- 9:59called VO3. Let's define VO3 for the listener.
- 10:03An AI engine that turns static images into moving
- 10:06clips. Simple enough, how do we actually use
- 10:08it? You take that full 12 -panel storyboard image.
- 10:11You upload that single image file directly into
- 10:13Google Flow. OK. Then you paste those 12 specific
- 10:16video prompts, the ones Claude wrote for us back
- 10:19in Stage 1. So we are giving Flow the static
- 10:21visual grid plus the text instructions of how
- 10:24things should move. Exactly. And you add a master
- 10:26instruction to the top of the prompt. You tell
- 10:28Flow, generate a scene using shots in the uploaded
- 10:31film storyboard sequence. The guide mentions
- 10:33adding one more strict constraint here. You have
- 10:36to explicitly tell the AI, no subtitles, and
- 10:39no music. Why do we need to specify that? We
- 10:41obviously want sound and text in our final video
- 10:43eventually, right? We do. But AI video generators
- 10:47have a really bad habit. If they try to generate
- 10:49audio or text alongside the video, it bakes it
- 10:52directly into the file. It's permanently attached
- 10:54to the visuals. Yes. AI -generated text often
- 10:57looks like alien gibberish. It hallucinates weird
- 11:00letters. AI -generated music might have tempo
- 11:03changes you hate. That makes sense. If that audio
- 11:06is baked into your raw video file, you cannot
- 11:08separate it later. You ruin your non -linear
- 11:11editing options. Blank audio and clean frames
- 11:14give you total control during the final edit.
- 11:16It beat. Exactly. Keep the raw footage as clean
- 11:18as possible so you hit generate and flow goes
- 11:21to work. It processes the grid. It does. And
- 11:24whoo! Imagine turning a flat 12 -panel grid into
- 11:2815 seconds of cinematic motion instantly. Yeah.
- 11:32It analyzes the spatial relationships in the
- 11:34static image. It calculates the temporal consistency
- 11:37required to make it move. It just blows my mind
- 11:39every time I see it work. It's pretty wild to
- 11:42think about. You are taking a flat static comic
- 11:46strip and getting a living, breathing movie out
- 11:48of it. Absolutely. It feels like something straight
- 11:51out of a sci -fi novel. But wait. We have 12
- 11:54distinct panels, and flow generates a 15 -second
- 11:58clip. That math is pretty tight. It is very tight,
- 12:01averaging just over one second per shot. Some
- 12:03of those shots must flash by incredibly quickly.
- 12:06They absolutely do. Some shots will feel rushed.
- 12:08Occasionally, the AI might even skip a panel
- 12:10entirely if the transition is too complex. Oh,
- 12:13really? Yeah, but that is OK. This initial generation
- 12:15is basically an animatic. It is a rough cut.
- 12:18It shows you the overall pacing and flow. So
- 12:21you watch this rough 15 second clip, what are
- 12:23we looking for? You are checking for temporal
- 12:26consistency. Do the warm orange tones of the
- 12:29volcano ash stay consistent into the final gas
- 12:32station scene? Does the camera pan smoothly?
- 12:35If it feels too fast, or if a transition is jarring,
- 12:39how do we fix it? We aren't in a traditional
- 12:41editing timeline here. Flow has a feature called
- 12:44the describe your edits box. It uses natural
- 12:46language processing to adjust the timeline. You
- 12:49literally just type smooth the transition between
- 12:52each shot, or you type maintain consistent pacing
- 12:55throughout the video. You just ask it nicely
- 12:58to fix the edit. Exactly. The engine recalculates
- 13:01the latent space between the frames. It generates
- 13:04new transitional frames to smooth out the motion.
- 13:07Wow. It adjusts the pacing without changing the
- 13:10core visual assets you establish. That is incredible
- 13:13control. You are directing the edit with text.
- 13:16And because you did the hard work in stage one
- 13:18and two. Because you built that perfect design
- 13:21sheet. Right. The final motion clip maintains
- 13:23its integrity. The characters look right. The
- 13:25lighting matches. It feels intentional. The Claude
- 13:28to chat GPT to flow pipeline is undeniably powerful.
- 13:32But let's look at the bigger picture. Not everyone
- 13:35uses those specific tools. Maybe they don't have
- 13:37access. Maybe the subscription costs for three
- 13:39different premium AI services are just too high.
- 13:42That is a very valid concern. Generating images
- 13:45and video at scale gets expensive quickly. The
- 13:48beauty of this specific framework is its modularity.
- 13:51The tools are completely swappable. Let's dig
- 13:55into the tool built. What are the alternatives
- 13:58for stage one, the prompt writing phase? If you
- 14:01don't want to use Claude, Gemini, is a fantastic
- 14:04alternative. Google's Gemini has a very capable
- 14:08free tier. It handles large context windows beautifully,
- 14:11which is great for building those skills. What
- 14:13about Grok? Grok is another great option for
- 14:16the text phase. It is incredibly fast. It also
- 14:18has fewer safety restrictions, which can be helpful
- 14:21if your story involves action or conflict that
- 14:23other AIs might mistakenly flag as inappropriate.
- 14:27What about stage two, generating the static images?
- 14:30We need tools that excel at building consistent
- 14:33design sheets. Ideogram is a phenomenal alternative
- 14:36to chat GPT and Dale E3. Why ideogram specifically?
- 14:40Two reasons. First, it is currently the best
- 14:42in the market at rendering visible text accurately.
- 14:45Oh, nice. If your scene requires a neon sign
- 14:47or a newspaper headline, ideogram nails it. Second,
- 14:50it has unique built -in tools for maintaining
- 14:52strict character consistency across different
- 14:54prompts. Any other image alternatives? Leonardo
- 14:57AI. It has an amazing free tier. It gives you
- 14:59an incredible amount of granular control. over
- 15:02the artistic style. If you want a highly stylized,
- 15:05painted look rather than photorealism, Leonardo
- 15:08is brilliant. Now for the heavy hitter. Stage
- 15:11three, the video generation. This is almost always
- 15:15the most computationally expensive step. What
- 15:18can we use instead of Google Flow and VO3? Kling
- 15:213 .0 is a massive contender right now. I have
- 15:24seen a lot of clips from Kling online. This looks
- 15:26very cinematic. It is. Kling 3 .0 handles physics
- 15:30simulations remarkably well. Water splashing,
- 15:34smoke billowing. It's often cheaper than enterprise
- 15:36solutions, and the quality is stunning. What
- 15:39else is out there for video? Pix4C1 is another
- 15:42strong option. It is specifically optimized for
- 15:44taking storyboard reference panels and translating
- 15:46them into consistent motion. Good to know. And
- 15:48if you are on a strict $0 budget, Hylua AI is
- 15:52perfect for completely free experimentation.
- 15:54So you have all these clips, but we still need
- 15:56to put it all together. We need to add the music
- 15:58and subtitles we avoided earlier. For final editing,
- 16:02CapCut is the dominant choice for creators. Right.
- 16:04It's free, incredibly intuitive, and there is
- 16:07no watermark if you use the desktop version.
- 16:10If you want Hollywood level control, DaVinci
- 16:12Resolve offers professional color grading. It
- 16:15is a massive complex program, but the base version
- 16:18is entirely free. And obviously Adobe Premiere
- 16:21if you are already in that ecosystem. But let
- 16:23me ask you this. If we swap out the software,
- 16:26if we use Gemini instead of Claude and Kling
- 16:29instead of Flow, does the quality of this specific
- 16:32sequence degrade? Does the magic disappear? Not
- 16:35at all. The discipline of the sequence is what
- 16:37creates the quality. It is absolutely not about
- 16:39the specific brand of AI. It's the architecture
- 16:42of the workflow. Precisely. The AI models will
- 16:45change. A new tool will launch next week that
- 16:47makes VO3 look old. But the logic of the pipeline
- 16:50remains. Right. You must build the textual blueprint
- 16:53first. You must lock in the static image second.
- 16:56You only animate as the absolute final step.
- 16:59The sequence is universal, even if the exact
- 17:01software you decide to use changes. Two -sex
- 17:04silence. That is the big takeaway here. Building
- 17:06an AI video from scratch feels like magic. When
- 17:10you watch a finished high -quality AI film, it
- 17:12looks like a miracle of prompting, but it isn't
- 17:15magical prompting at all. No, it's just a structured
- 17:17process. It is about having a clear, methodical
- 17:20sequence. You start with a simple story. You
- 17:23let the text tool structure your prompts. You
- 17:27build your static visuals to anchor the AI. Only
- 17:31then do you hit the animate button. Each tool
- 17:34handles one specific job. You isolate the variables.
- 17:38You fix problems while they are still flat images.
- 17:41Because fixing a moving target is nearly impossible.
- 17:43Exactly. You become a director, carefully setting
- 17:45up the shot. You stop being a gambler just pulling
- 17:48the slot machine lever. It changes everything.
- 17:50Yeah. I want to leave everyone with a final thought
- 17:52to chew on today. We have been talking about
- 17:54this specific AI video workflow. But think about
- 17:58your own workflows outside of video creation.
- 18:00Oh, it applies to almost everything. It really
- 18:02does. Think about analytical business projects
- 18:05or writing code or even designing a presentation.
- 18:08This methodology building the design sheet first
- 18:10is universal. How often do we rush to the final
- 18:14moving product? We jump straight into the execution
- 18:16phase. blindly because it feels productive, we
- 18:20try to animate our ideas before we have actually
- 18:23designed them. We skip the blueprint because
- 18:25we want to see the house. Exactly. And then we
- 18:27spend twice as long fixing structural errors
- 18:29that never should have happened. Right. Next
- 18:32time you start a complex new project, stop and
- 18:35ask yourself, did I build my static design sheet
- 18:38first or am I just hoping the sequence works
- 18:41itself out? Thank you for joining The Deep Dive.
- 18:44We will see you next time.