Latest / AI Fire Daily / #526 Neil: AI Video Creation With Claude Fable 5 And Google Flow
Transcript
- 0:00You know, that feeling when you make an AI video.
- 0:04You spend hours crafting the perfect initial
- 0:08text prompt. You finally generate this absolutely
- 0:10beautiful first video clip. Beat. It looks exactly
- 0:14like the vision in your head. Yeah, I mean, the
- 0:16lighting is just totally gorgeous. Right, and
- 0:18the physical motion is incredibly smooth, but
- 0:20then the second clip completely ruins the entire
- 0:23sequence. Oh, absolutely. The main character's
- 0:25face just morphs entirely into someone else.
- 0:27Or their red jacket mysteriously becomes blue
- 0:29denim. The setting randomly jumps around without
- 0:32any logical reason. A quiet coffee shop suddenly
- 0:35looks like a crowded spaceship. It is, well,
- 0:39a deeply universal frustration for many creators.
- 0:41You just feel like you have absolutely no creative
- 0:43control. Exactly. Beat. Let's unpack this. Welcome
- 0:47to our deep dive into continuous storytelling
- 0:49today. We are going to master a very specific
- 0:52workflow. We will solve this exact continuity
- 0:54problem step by step. And we are pairing two
- 0:57extremely powerful tools to achieve this. We
- 0:59are using ClaudeFable 5 and Google Flow. We will
- 1:02break down a clear, highly repeatable methodology
- 1:05today. This method does not rely on getting lucky
- 1:09prompt generations. It keeps your characters
- 1:11perfectly consistent across long video sequences.
- 1:15It stops your story from randomly breaking apart
- 1:17completely. Two -sec silence. Before we generate
- 1:21any video, context really matters here. We must
- 1:24understand why we're using these specific tools
- 1:27right now. Right, so Cloud Fable 5 actually came
- 1:30back online recently. It was blocked briefly
- 1:32due to strict U .S. export rule. Yeah, Andropic
- 1:35had to take it down for a short time. Now it
- 1:37operates with a brand new safety layer. Meaning
- 1:39filters that stop the AI from making harmful
- 1:41or banned stuff. Exactly. It was a brief hiccup,
- 1:44but the new version feels incredibly stable.
- 1:47It is highly reliable for complex, multi -step
- 1:50reasoning tasks. We use Fable 5 specifically
- 1:53for the prompt writing tasks. It handles extremely
- 1:55long and layered prompts quite beautifully. It
- 1:58can balance specific character details and complex
- 2:00scene logic. Best of all, it does not drop important
- 2:03details halfway through. Then we pair it directly
- 2:05with Google Flow. Flow runs on the powerful Veo
- 2:08video generation engine. Flow is crucial here
- 2:10because it takes reference images perfectly.
- 2:13It actually uses those reference images as core
- 2:16storytelling ingredients. Most platforms struggle
- 2:19to lock in specific visual character traits.
- 2:22Flow actually anchors the generated video directly
- 2:24to your uploaded image. The workflow is honestly
- 2:28quite simple. You write the complex prompt inside
- 2:30of Fable 5 first, then you pass it directly into
- 2:34Google Flows visual interface. There is no middleman
- 2:37tool required in this entire workflow. I guess
- 2:40what I'm wondering is why the handoff? It feels
- 2:45like a lot of extra work. Why not just generate
- 2:47the video natively inside Fable or do the prompt
- 2:50writing directly inside of Flow? You would think
- 2:53that would be easier, right? But it is fundamentally
- 2:55a left brain, right brain problem. Flow has this
- 2:58hyper -realistic visual engine, but its underlying
- 3:02narrative memory is basically a goldfish. So
- 3:04it forgets that your character's jacket has a
- 3:06specific zipper. Exactly. Fable, on the other
- 3:08hand, has superior logic for tracking very complex
- 3:11details over time, but it doesn't actually generate
- 3:13the video. You really need Fable's brain and
- 3:15Flow's visual paintbrush. Normal AI video generation
- 3:18is a highly chaotic process. It feels a lot like
- 3:21navigating a very chaotic dream. Faces constantly
- 3:25morph and shift without any warning whatsoever.
- 3:28But using Fable 5 and Flow together is entirely
- 3:31different. It's like stacking Lego blocks of
- 3:33data. It keeps the underlying reality completely
- 3:37solid and grounded, beat. So Fable writes the
- 3:40exact instructions and Flow paints the moving
- 3:43picture. Exactly. You need both to keep the illusion
- 3:45alive. Two -sex silence. Okay, so Fable is the
- 3:49brain and Flow is the canvas. But even the best
- 3:52painter needs a strong visual reference. If we
- 3:55just toss our character into a scene blindly.
- 3:57The AI is going to improvise her face differently
- 3:59every time. Right. So how do we lock in her DNA
- 4:02before we yell action? We need a very strong
- 4:04visual foundation to start with. We begin by
- 4:06generating a master character sheet first. This
- 4:09happens first inside of the Fable 5 chat window.
- 4:11You describe your character with extreme visual
- 4:14and physical detail. Let's say you ask for an
- 4:16Asian woman in her late 20s. She needs shoulder
- 4:18length, black hair, and very sharp facial features.
- 4:21Yeah, and she wears a dark red bomber jacket
- 4:24and cargo pants. She has worn white sneakers
- 4:27and a small black backpack. You must request
- 4:30multiple camera angles in this initial prompt.
- 4:32You need front views, profile views, and detailed
- 4:35close -ups. And here is the kicker. You must
- 4:37specify a completely neutral background environment.
- 4:41You paste this detailed prompt into Google Flow
- 4:43next. You select the Nano Banana Pro image model
- 4:47for this. You set it to high quality for a sharp
- 4:50reference image. Flow returns a sheet with several
- 4:53angles of your character. Then you jump back
- 4:55over to the Fable 5 interface. You stay in the
- 4:57exact same chat window session. Fable remembers
- 5:01every single detail about her specific outfit.
- 5:03You just ask it for the first frame of the scene.
- 5:05A Chinatown dumpling shop on a very warm evening.
- 5:08Steam is rising slowly off the hot steaming food
- 5:11trays. Warm light is coming directly through
- 5:13the front shop window. You generate the still
- 5:15frame inside of Google Flow next. You attach
- 5:17the character sheet as your core reference image.
- 5:20Flow gives you the first still frame of the scene.
- 5:23It shows the character and the beautiful setting
- 5:25perfectly together. Yeah. I have to admit something
- 5:30here. I still wrestle with prompt drift myself.
- 5:33Oh, it happens to everyone. I lose track of character
- 5:35outfits all the time. I'll get three scenes in
- 5:37and suddenly my hero is wearing a completely
- 5:40different shirt. It drives me crazy. That is
- 5:42incredibly common and it is exactly why the neutral
- 5:44background setting matters here. I wanted to
- 5:46ask about that actually. Why not generate her
- 5:48in the shop right away? Why take the extra step
- 5:51of a blank background? Because we have to think
- 5:54about how AI sees images. When AI models generate
- 5:58video, they are constantly looking for visual
- 6:00patterns. They reconstruct images from pure digital
- 6:04noise. If you put a character in a red jacket
- 6:06inside a red room, the model's attention mechanism
- 6:09gets completely confused. And an attention mechanism
- 6:12is basically how the AI decides which specific
- 6:15pixels matter most. Right. So the math literally
- 6:18cannot tell where the jacket ends and the wall
- 6:20begins. The pixels just bleed together entirely.
- 6:23That is why we have to isolate her on a neutral
- 6:25background frame. Mm -hmm. It forces the AI to
- 6:28learn her specific edges. It learns the exact
- 6:31shape of her face without any shadows confusing
- 6:33the data. It is like taking sterile actor mug
- 6:36shots. You capture them perfectly before dropping
- 6:39them onto a chaotic movie set. That is a perfect
- 6:42way to look at it. A clean reference isolates
- 6:45her core visual identity completely. You need
- 6:48a clean master blueprint before dropping characters
- 6:50into a messy scene. Two -sec silence. All right,
- 6:54we have our hero perfectly framed right now.
- 6:56We have our atmospheric setting established beautifully.
- 7:00But a good story desperately needs some real
- 7:03narrative conflict. It requires physical motion
- 7:05and actual dramatic action. Yeah, you need something
- 7:08to actually happen. Right. So how do we introduce
- 7:10action into this quiet dumpling shop scene? We
- 7:13need to create some bad guys next. We go back
- 7:16to Fable 5 for a new prompt. We ask for a brand
- 7:18new character reference sheet. We need three
- 7:21thugs in their 30s for this scene. They wear
- 7:23black leather jackets, torn jeans, and heavy
- 7:25boots. But we can't just leave it at that, right?
- 7:27No. We have to give each one a very distinct
- 7:30visual detail. This ensures they are very easy
- 7:33to tell apart later. One has a dark neck tattoo,
- 7:36clearly visible. One wears a baseball cap pulled
- 7:39down low. One is noticeably taller than the other
- 7:41two men. We generate this sheet using Nano Banana
- 7:44Pro in Flow. Now we have two completely distinct
- 7:47character reference sheets. We have one for our
- 7:50hero and one for the thugs. Now we rate the prompt
- 7:52for the actual video clip. We use Gemini OmniFlash
- 7:56for this specific part. The prompt needs a start,
- 7:58a change, and an end. The woman stands quietly
- 8:01inside the warm dumpling shop. The three thugs
- 8:04suddenly burst through the front shop door. She
- 8:06quickly turns to run toward the back kitchen
- 8:08area. Fable returns a prompt that lays out this
- 8:11sequence perfectly. We switch Google Flow into
- 8:14video generation mode next. We set the duration
- 8:16to exactly 10 seconds. We attach the hero reference
- 8:19sheet and the thug reference sheet. We also attach
- 8:22that very first frame image we created earlier.
- 8:24Flow generates a 10 second clip of this exact,
- 8:27intense action. Beat. Whoa! Imagine perfectly
- 8:31directing three distinct, consistent characters
- 8:34bursting into a scene, all without ever touching
- 8:37a real camera. It is pretty wild. What's fascinating
- 8:40here is how the AI handles those traits. We give
- 8:43them a tattoo, a hat, and a height difference.
- 8:45Those distinct details literally act as mathematical
- 8:48anchor points. They stop the characters from
- 8:51blending together completely during fast motion.
- 8:53Exactly. Without the tattoo or the cap, they
- 8:56melt together. The AI would turn them into one
- 8:58generic leather blob. Because it just sees a
- 9:01mass of dark pixels moving fast. Right, and it
- 9:04tries to average them out. The distinct traits
- 9:07force the attention mechanism to track three
- 9:09separate physical objects. That makes total sense.
- 9:12But let me ask you this. Why limit the tromp
- 9:14to exactly 10 seconds if we have all this great
- 9:17reference material? Why not just generate the
- 9:19whole two -minute chase at once? It comes down
- 9:22to how these video models actually simulate reality.
- 9:25They don't actually understand physics or object
- 9:28permanence at all. They're just predicting the
- 9:31next logical state of pixels. Over long stretches,
- 9:34that prediction math gets very muddy. The models
- 9:36literally lose track of gravity. So a 10 second
- 9:40clip is the absolute limit for physical consistency.
- 9:42Exactly. It keeps the heavy action manageable
- 9:45and highly predictable. If you push for a full
- 9:47minute, the characters will start floating. or
- 9:51walking straight through solid walls. The math
- 9:53just degrades too far from the original starting
- 9:56point. Short bites keep the AI from hallucinating,
- 9:59making it easier to stitch together later. Neat.
- 10:02We will explore how to perfectly extend this
- 10:05chase sequence. Right after, a quick word from
- 10:09our sponsors. Stick around. Okay, so we are back
- 10:12to our deep dive. The first 10 seconds of action
- 10:15are completely finished. Our main character is
- 10:17running toward the back kitchen area. The three
- 10:19thugs are bursting through the front door. The
- 10:20physics held up perfectly in that clip. But how
- 10:23do we continue the chase seamlessly from here?
- 10:25How do we stop the AI from completely forgetting
- 10:27the scene we just built? This is where we use
- 10:29the critical extension method. Gemini OmniFlash
- 10:32supports conversational multi -turn edits naturally.
- 10:36You go back to Fable 5 to write the continuation.
- 10:39But you absolutely must follow one crucial rule
- 10:41right here. You have to describe the exact last
- 10:44frame first. You can't just say what happens
- 10:46next immediately. No. You must say, the woman
- 10:49just ran into the back kitchen area. She is breathing
- 10:52hard with three thugs chasing behind her. You
- 10:55establish the exact physical reality of that
- 10:57final frame. Then you add the brand new action
- 11:00sequence next. A man on a motorcycle appears
- 11:03at the back door. He yells for her to quickly
- 11:06get on the bike. She jumps onto the back seat
- 11:08of the waiting motorcycle. It confirms the last
- 11:14state, then pushes the next action. If you skip
- 11:16confirming the last frame, it breaks completely.
- 11:19OmniFlash will completely lose track of the entire
- 11:21scene momentum, because remember, it lacks object
- 11:24permanence completely. By describing the last
- 11:27frame, you give it mathematical context. You
- 11:29are reestablishing the physical momentum of those
- 11:31pixels. It is a lot like a relay race, you know?
- 11:33Oh, that is a great analogy, actually. EI has
- 11:36to firmly grab the baton first. It must confirm
- 11:39the last frame to gain visual traction before
- 11:42it can... and went forward into the brand new
- 11:43action. You stay in the exact same video session
- 11:46in Flow. You paste the prompt from Fable into
- 11:49the edit box. You generate another 10 second
- 11:52segment of the ongoing chase. Flow picks up from
- 11:55the very last frame perfectly. It calculates
- 11:58the new motion vectors based on your text. It
- 12:00moves seamlessly into the new motorcycle escape
- 12:03scene. Then you must drag both clips into your
- 12:06video editor. You look closely at where the two
- 12:09clips actually join. You usually see a small
- 12:11jump right at the cut. The lighting might flicker
- 12:14or a hand might twitch. You trim a few frames
- 12:16off the start or end. This completely hides the
- 12:19seam between the two generated clips. It makes
- 12:21the action look entirely continuous and smooth.
- 12:24Wait, but if I'm already in a continuous chat
- 12:26session with OmniFlash, why do I have to manually
- 12:29reattach the character reference sheets repeatedly?
- 12:32Shouldn't the AI just remember what she looks
- 12:34like from the previous 10 seconds? You would
- 12:36definitely think so. But video generation models
- 12:39suffer from rapid visual memory degradation.
- 12:43Think of it kind of like making a photocopy of
- 12:45a photocopy. Each new generated frame becomes
- 12:48slightly less accurate than the last. Oh I see.
- 12:51The sharp edges of her face start to blur. Her
- 12:54red jacket starts to look a little faded. The
- 12:57AI gets lazy with the details over time. Exactly.
- 13:01Without the original anchor, the visual data
- 13:03degrades very quickly. Reattaching the original
- 13:06character sheet completely resets the AI's visual
- 13:08baseline. It forces the model to look back at
- 13:11the master blueprint. It recalculates her exact
- 13:14facial structure for the new clip. Even in a
- 13:16continuous chat, the AI needs the original photo
- 13:19to prevent drifting. Two sec silence. And I assume
- 13:23this applies to new props, too. If you want to
- 13:25add completely new things, like that motorcycle
- 13:27waiting in the back alley, let's make a new reference
- 13:30sheet for it first. Absolutely. You build it
- 13:32exactly like you build the character sheet. You
- 13:34prompt fable, generate a still image and flow,
- 13:36and attach it. Otherwise, the AI is just guessing
- 13:38what a motorcycle looks like, and it might give
- 13:40you a bicycle with a jet engine attached. Two
- 13:42sec silence. Let's zoom out from the specific
- 13:45timeline for a minute. We have talked a lot about
- 13:49the mechanics today. But let's look at the overarching
- 13:52philosophy of this entire system. It all comes
- 13:54down to a very strict, reliable loop. You write
- 13:57the highly detailed prompt inside of Fable 5.
- 14:00You generate the image or the video in Google
- 14:02Flow. You check the visual result carefully.
- 14:05Then you extend the scene by confirming the last
- 14:07frame. You push the new action forward. Then
- 14:10you join and trim the clips in your external
- 14:12editor. You repeat this loop until you have every
- 14:14story segment. So what does this all mean? The
- 14:17core lesson here is to start very small. Focus
- 14:20entirely on one single character at the very
- 14:23beginning. Build one perfect character sheet
- 14:25to practice this workflow. Pay attention to how
- 14:28the AI handles the neutral background. Generate
- 14:30one flawless 10 second video clip of simple action.
- 14:33Do one clean extension and one simple edit pass.
- 14:36Master this basic loop before building a massive
- 14:39action sequence. Pete, I invite you to try this
- 14:42workflow out tonight. Try building just one character
- 14:45sheet of your own design. See if you can maintain
- 14:48their face perfectly across two scenes. Try placing
- 14:51them in two completely different environmental
- 14:53background settings. You will completely change
- 14:55how you think about AI video. If this is the
- 14:58incredible storytelling fidelity we can achieve
- 15:00right now, just by manually stitching together
- 15:03short 10 -second AI clips, what happens to the
- 15:06entire film industry very soon? When these video
- 15:09models can natively hold character consistency
- 15:12perfectly. For a two -hour feature film without
- 15:14any human trimming required. That is a truly
- 15:16staggering thought to leave you with. Thank you
- 15:19for exploring this deep dive with us today. Keep
- 15:21experimenting and pushing the boundaries of these
- 15:23amazing tools. We will see you back here next
- 15:25time.