Speaker A
Heat. Heat. If you keep playing on this knee, you may not walk properly again.
Higgsfield AI creator Adil walks through making a 72-hour AI commercial with Seedance 2.5, covering character assets, prompts, scene blocks and editing.
Key Takeaways
- A structured four-beat story plus block-based prompting makes multi-character AI commercials manageable.
- Consistent character sheets and named asset elements are essential for keeping faces and outfits stable across shots.
- Using an LLM with a reusable directing skill speeds up shot planning and prompt writing.
- Multiple takes and selective editing compensate for model inconsistencies in motion and framing.
What the video covers
- The video shows how to build a multi-hero AI commercial using a four-beat story structure: setup, problem, choice, and resolution.
- Characters are created in Soul Cinema using a clean character sheet (gray background, body views, close-up face) to keep them consistent across scenes.
- Logos and outfits are generated in GPT Image and Cream, with edit prompts written by Claude and applied to the character's face-consistent base image.
- Claude is loaded with a custom 'directing workflow' skill that converts scene descriptions into shot plans, camera movement, and performance direction.
- Assets such as locations, props (knee brace), and character sheets are saved as named elements so the right asset attaches to each prompt.
- Scenes are generated in Cinema Studio using Seedance 2.5, with multiple takes per prompt and the best moments selected and combined in editing.
- Color looks are controlled through Cinema Studio's hex color presets; the creator rejects a purple 'Twilight Fable' preset in favor of 'Morning After Rain' for a neutral palette.
- Fixes include adding wide establishing shots when a close-up prevents the model from showing the full body, such as a character standing at a table.
- The story is split into five blocks, with one prompt potentially covering several scenes, and a second hero's story runs in parallel with the goalkeeper's.
- The video closes with a cost breakdown including failed takes, and points viewers to linked prompts, assets, and the blog post.
Chapters
- 00:00Intro and Story Structure
- 03:00Character and Asset Creation
- 07:00Logo and Outfit Design
- 09:00Directing Skill Setup
- 11:00Setup Block: Clinic and Home
- 17:00Problem Block: Training Field and Color Grading
- 22:00Choice Block: Changing Room and Brace
- 30:00Second Hero Story and Dance Battle
- 40:00Resolution and Final Cut
Full Transcript — Download SRT & Markdown
Speaker A
Don't tie your life to the street. I want to take care of you. I'll take care of myself.
Speaker A
Heat up here. Hi, I'm Adil. You guys often ask me how to turn your own characters and clothes into a commercial. It's three simple steps: assets, scenes, and editing. This story jumps between three heroes in one ad. So I built it on my four beat
Speaker A
script. Setup, problem, choice, and resolution. Every scene I generate gets one of those jobs that decides how I split the story into blocks and later how I cut them together. Take the goalkeeper. His fall during training is the problem, and the save in the final
Speaker A
match is their resolution. So there are two separate blocks. And in the edit, the save sits right next to the other two heroes resolutions. Every prompt and asset, including the scenes I skip, is in the blog post below. And at the end,
Speaker A
I'll show you what this actually cost, including the takes that didn't work. Now, before I generate anything, I turn the story into a short list that tells me which characters, clothes, and locations I actually need. For every shot, I describe it to Claude. Claude
Speaker A
writes the prompt, and I generated in Cinema Studio with my assets. Everything lives in one project, a folder for each hero. Once an asset is done, I save it in Hixold as an element with an at name, and I give Clot the same image with the
Speaker A
same name. That way when I paste Claus prompt back into Higseld, the right element attaches itself. Every character gets the same clean sheet, a gray background, body views with a head cropped out, and a big close-up of the
Speaker A
face. That's what keeps him looking the same in every scene. Goalkeeper first. I'm asking Claude for an ordinary working guy. Dark curls, a warm face, real skin that holds up in a close-up.
Speaker A
And he's a goalkeeper, so even in regular shirt, he has to look like an athlete. Now, I'm taking Claude's prompt into Soul Cinema and generating four options at once. for people. That's my go-to image model. It keeps real skin
Speaker A
textures so faces don't turn into a plastic AI slop. Now, the first one smiles a little too much for me. The second one feels too stern. He looks like he's about to give me a parking ticket. The third is closer, but the
Speaker A
smile still feels posed, but the fourth one, I like this guy. He looks like a real person, not a model. Let's go with him. And let me know in the comments which one you would go with. Now, before
Speaker A
I dress him, we need a logo. Simple enough to spot on a shirt or a shoe.
Speaker A
Nothing fancy. So, I'm switching to GPT image. It's the best at clean flat graphics. I'm generating four options.
Speaker A
The loops feel interesting, but the double outline on someone's chest is just noise. But the C, yeah, it's one wide open shape. So, even small on a shirt, you can tell what it is. I'm saving it as an asset so the model
Speaker A
doesn't invent a new logo for every outfit. Now, for the kit, we'll do it once in detail. Then, it's the exact same move for everyone else. Claude gets his character sheet and the outfit sheet and writes the edit prompt. I run it in
Speaker A
Cream inside Cinema Studio. Now, Cadream is great at edits. He changes the clothes and keeps his face exactly the same. And there he is. Same face, but now in the green kit. I'm checking the crest, panels, number, and logo against
Speaker A
the outfit sheet. Now, quick note, these clothes are generated for your own product. Just shoot it front and back.
Speaker A
That's how the model gets the cut, the print, and the logo exactly right. All right. He also needs a version with a knee brace because he cuts it off later.
Speaker A
That means a clean prop sheet of the brace. Then the kit with it on. And here's the brace version. Same kit and face with the brace on the right knee.
Speaker A
I'll bring in everyone else's looks when a scene needs them. And you can grab all those prompts from the link in the description. The assets are ready. Now the scenes. And to write those prompts, Claude needs to know how I direct. So I
Speaker A
give it my own directing skill. A set of instructions you load once and Claude follows them every time. This one's called directing workflow. It turns a scene into a shot plan. It's linked below. Loading it takes a few seconds.
Speaker A
Customize skills. upload the file and switch it on. Then I just mention it in my request. From now on, I work like this. I split every story into blocks.
Speaker A
One block is one prompt, and one prompt can hold several scenes. I run it a few times, pull the best moments from different takes, and fix whatever breaks. First up, an injured goalkeeper with one last big match and a knee that
Speaker A
might not hold. I build a story from five blocks. The clinic, his home, and the street are the setup. Training is the problem. The tunnel and changing room are his choice. and the match is the resolution. Block one is the setup
Speaker A
that establishes the main character and his reality. Two scenes and one prompt. At the clinic, the doctor tells him how bad his knee is. And at home, he pulls the praise on. So, I'm bringing in all the assets for the scene, location
Speaker A
assets for the clinic and the flat, his character sheets, and his everyday clothes in a home outfit. And of course, the brace. Then, I describe what I want in the first scene to Claude. That's all I type. The skill does the rest. It
Speaker A
breaks the scene into shots, decides how the camera moves, and even tells the model how he should look, lost and empty. Then I paste the prompt into Cinema Studio and pick CS2.5. I'm generating two takes. Okay. In both,
Speaker A
he's slumped on the table, practically lying down. I want him standing against it, and the prompt says standing, but the shot starts so close that the model never sees his whole body. The home half is fine in both takes. The acting works,
Speaker A
so I'm only fixing the clinic. For that, I'm adding a wide shot at the start, so the model sees him standing at the table before we move in. That makes four shots. And there we go. He's upright in
Speaker A
both, and he stays that way. Now, I'm taking the best of the start and the end. The clinic from one take, his home from another one. Let's watch the clinic and his home cut together.
Speaker A
The damage to the ligament is serious. If you keep playing on this knee, you may not walk properly again.
Speaker A
Before he gets back to training, I wanted one more scene to make his setup feel real.
Speaker A
He walks past a mural of himself under a bridge just as someone sprays retire old man across it. That's block two, one scene. The mural is part of the location. I'm starting with an empty street under a bridge. Then a serum edit
Speaker A
paints the mural onto the pier and hangs an old lamp right above it. So the light on the mural matches the street. In the request, the timing does the work. The car passes, the mural shows up behind it, and the vandal is just finishing up
Speaker A
the tag. This one worked on the first try. I'm taking the walk and the reveal from one take and his face from the other.
Speaker A
Next comes the problem, the obstacles in his way. He tries to come back to his training, but the knee says no. First though, he needs somewhere to train. I build a field at a 3/4 angle, so the model gets the depth of the pitch and a
Speaker A
clear line to run. For lighting, it's blue hour. Cool, but bright enough to follow the action. Claw's prompt goes into Cinema Studio image soul cinema.
Speaker A
And before I hit generate, I pick the color. So, Cinema Studio has built-in hex color presets. And you choose them up front. First test, Twilight Fable.
Speaker A
Let's make a batch of four images. Way too purple. In combination with my other shots, it would look like it's from another film, and I'm losing half the pitch in the shadows. I want it dark, but I still need to read the space. Same
Speaker A
prompt, but I'm switching the hex preset to the morning after rain and running for more. Oh yeah, the morning after rain works. Colder sky, no purple cast, more neutral shadows. This is the one I'm keeping. The field and the look are
Speaker A
locked. Block three. For this beat, I want a training sequence, a warm-up, then the drill, both in one prompt. He sprints to the cone, turns, and his knee gives out. That turn is the part I need to nail. two training sheets and the
Speaker A
brace stays on his right knee in both. I start with the warm-up. First try, the simple way. His character sheet, the field, and some words. No. Okay. He jogs and that's it. No rhythm, no knee lift.
Speaker A
It looks like a guy who's late for the bus. This is where I got stuck. I burned a bunch of credits trying to describe this warm-up in words and it never moved the way I wanted. So, I made it easy. I
Speaker A
filmed the movement myself. Watch the right knee. It comes up high on every other stride. And that's much easier to show in a clip than explain it with a natural warm-up. And here's how the clip gets in. In Claude, I'm attaching the
Speaker A
clip to my message together with his character sheet and the field. And I'm telling Claude what this clip is for.
Speaker A
Only the movement, the rhythm, and the knee coming up on every other stride. Claude writes me a full prompt that takes the motion from the clip and everything else from my assets. Then in Cinema Studio, I'm picking Cance 2.5.
Speaker A
Adding the clip next to the images as a video asset and pasting in the prompt there. So the clip is called at video 1.
Speaker A
Here's a warm-up I went with. And there's the knee lift. I couldn't get with words. Now to the second part of the blog, the fall. I wanted to read as the knee giving out, not a trip. So I
Speaker A
spell out the order. He's pressed to the cone, plants his right foot, starts to turn, and the knee gives way. He drops onto his left side, rolls onto his back and grabs his right knee, then a wide from the bench. The coach just stands
Speaker A
there with his arms folded and watches. Two more things in this prompt. For the acting, the directing workflow skill scripts his face beat by beat, lips pressed, jaw clenched, breathing hard through his nose. And for the sweaty look, just one line, sweat at the
Speaker A
hairline. This one's 26 seconds and nine shots. So, I'm not going to render it in full quality straight away. I'm making 480p drafts and Cents cheaper, faster, and they're good enough to judge the movement and the framing.
Speaker A
Okay, the knee lift from my clip made it in and the turn leads into the fall. So, the training scene comes from three takes, the sprint and the fall from one, the turn and the cone from another and
Speaker A
the coach from a third. With drafts, you only finalize the takes that go into the edit in 1080p. The warm-up comes from another take, so I stitch it together with a drill in the edit. Take a look.
Speaker A
Now, his knee gives way right as he turns. My own clip gave me the movement here. So, if you need a specific move, film it yourself first. It's cheaper than running take after take and hoping the model gets it. On to the choice.
Speaker A
Block four. Two scenes again. In the tunnel, the coach tells him management wants to replace him. That outside pressure pushes him. Then in the changing room, he cuts off the knee brace. That's his choice. For the tunnel conversation, they need space from the
Speaker A
walls so he can finish his line and walk out. In the changing room, it's a closeup. He cuts the brace and the camera pushes in on his face. Both scenes go into one request. Two takes this time. He's hugging the wall, but I
Speaker A
asked for both men in the middle. The changing room already works. He cuts the brace and pulls it off. that leaves the tunnel. I doubt more words would fix it, so I'll just draw it to save some time
Speaker A
and credits. Don't judge my drawing. Honestly, this is my best work. So, blue is the keeper, yellow is the coach. That just tells the model where they stand and which way they face. Now, with the prompt from Claude and the added sketch,
Speaker A
I run it again. And this time, both men are right in the middle of the tunnel.
Speaker A
So, now I can mix the shots from different batches. The tunnel from this one, him cutting the brace from the second, and the final look from the last one. Here's the mix.
Speaker A
Okay, on to the final part. The resolution comes in block five, his comeback after the fall. The walk out and the match share one prompt. We start behind him, walking out of the tunnel, the Dino one on his back, then the play,
Speaker A
the shot, the save, the counterattack, and his reaction. First, the bricks. Anything recurring more than once gets its own acid. His team, the opponents, the stadium, the ball, because with pure text to video, the kids and the players
Speaker A
would change every take. So, I gave them their own sheets. The crowd is already in the stadium asset, so the model doesn't invent people from scratch. And he's in his match kit, so no knee brace this time. Here's a request. It's quite
Speaker A
a detailed one. That's a lot of story for 22 seconds. Let's see what we got.
Speaker A
The walk out is spoton. The silhouette pagina one on the back. stewards at the tunnel. The kids hold every take, too.
Speaker A
Blue for home and red and black for the opponents. Now, the safe. He flies top corner, hugs the ball, grass sprays. I mean, I could watch that on a loop. So, the match comes from four takes of this
Speaker A
prompt. The walk out from one, the face, the shot, and the save from another, and the long kick, and his reaction from two more. I end on his reaction. He walks up, smiles, and raises one finger. His smile lets all that tension go.
Speaker A
One hero down, two to go. Same for be script for both of them. So, let's pick up the pace. Our next hero is a ballet dancer who wants to do hip-hop and her family doesn't get it. Her story takes
Speaker A
five blocks with the same four steps. It starts with the match cut and the dinner. The locker room is her problem.
Speaker A
practice in the argument with her dad or her choice. And it all ends at the dance battle. From here on out, I'm not walking through every asset step by step. You know the workflow. I'm going to stick to what's new in each scene.
Speaker A
Her setup is bulk wine and it opens on a match cut. She closes her eyes in the ballet studio, then opens them on the bridge. Her face stays in the same spot in the frame and only her clothes and
Speaker A
surroundings change. Then we cut to dinner. She sits next to her dad. He keeps eating. Her mom glances at her once and nobody says a word. I want you to feel how uncomfortable she is. The asses go in as usual. Her three looks,
Speaker A
both parents, and all three locations. I ask cloth for both scenes together. Here's what I got. The match got worse here, but the dinner scene, it's too wide. I'm watching the whole take care, not her. And I also want to put both
Speaker A
parents on the other side of the table to highlight the emotional distance between them. So from the steak, I'm keeping the match cut. So it's just the push in and the mother's position. I'm changing. Start on her empty chair. Push
Speaker A
in as she enters and land right on her face, keeping her dad at the edge.
Speaker A
Perfect. Once the camera knows where to go, I get the exact cut I want. I'm keeping two takes, one for the match cut and one for the dinner. In the edit, her setup looks just like this.
Speaker A
As for her problem, her parents are already against her dancing, but I wanted to push it further. So, block two is a single scene, the locker room, where she reads the negative comments under her dancing videos. The trick is
Speaker A
keeping her dance clip on the phone screen with a locker room around her. Here's the move. First, I generated the vertical clip of her dancing. Then, I uploaded it to Claude along with a phone asset and everything else we need for
Speaker A
the scene and write the following request. Here's the first take. The dance stays inside the display and the comments are readable. That's what she reacts to. I keep the close-up on the screen, cut to her face, and trim that
Speaker A
long pull back at the end. I run it a few more times and build a locker room scene from five passes, one take from each. The girl leaving, the phone, the comments, her lying down, and the wide shot from the other one. Put together,
Speaker A
it plays like this. Her choice comes in two halves. In the first, she keeps practice despite the comments. Book three covers that. The bridge and the studio, an outfit for each in one prompt. And now she has to
Speaker A
really dance. First test, black streetear on the bridge, dancing toward the camera. I'm keeping the request simple. She moves like a robot. And then she just starts jogging. I asked to mix walking with dancing. What I wanted is
Speaker A
one move flowing into the next. Shift the weight, turn, drop low, and keep going. On my first try, I didn't have a dance skill. So, I fed Claude videos of my friends dancing to break down the footwork, weight shifts, and arms and
Speaker A
turn that into a dance direction skill. It loads the same way as the first one.
Speaker A
Now, instead of just typing make her dance, I can prompt actual steps. Two side steps, a quarter turn, a low drop.
Speaker A
Those are easy to check. Eight shots in all, four for each location. With the skill, I try it twice. Here they are side by side. Without the skill and with it. Now that's dancing on the bridge and in the studio. She moves through the
Speaker A
steps instead of breaking into a run. That gives me the practice. She keeps dancing no matter what they wrote.
Speaker A
Her story needs two more parts. She stands up to her dad and then steps into the dance battle. The argument comes first. One scene and it finishes the choice part. Her dad follows her down the hall and she answers firmly before
Speaker A
closing the door. So, I'm describing that to Claude. And this is what comes back.
Speaker A
Don't tie your life to the street. I'll take care of myself. It's clear who's speaking. She lands her answer and the door shuts after it. I just trim the walk and keep the stake.
Speaker A
I'm leaving out a few scenes like her grabbing the sneakers. You'll find every one of them with the prompts and assets again at the link below. The battle is her resolution. Block five with three scenes. her footwork, the celebration,
Speaker A
and her dad watching from the crowd. She needs to own the center of the circle with a judge and opponent framing the edge. For the first take, I even tried having the opponent step in while his crew pulls him back. I'm attaching all
Speaker A
the assets for the scene. The venue that already brings the crowd, her opponents, and the judge's character sheets, and her final look, and obviously my request to clot. Eyes on the opponent, bro. It's not even his turn. Actually, I don't
Speaker A
think I like him coming into the circle. My eyes go to him when I want to be washing her feet. That interruption was on me. I thought it would add some tension, so I'm taking his entrance out.
Speaker A
He stays behind the line, and I keep the celebration and her dad's reaction. One change to clot. The center stays clear, her feet are visible, and he never crosses the line. I run the new version twice. The first low shot comes from one
Speaker A
take, and other two shots come from the second one. The crowd lifts her, but her father is there, too. His reaction wraps up the family story.
Speaker A
Without it, the argument would just end at the closed door. Last up, the school boy. The quiet kid gets called a nerd.
Speaker A
Again, it's five blocks, and by now you know the four steps. School and home, the note, the jersey walk, and the tournament. His setup shows two sides of his life, so that's two blocks, one scene each. Block one is school, where
Speaker A
he keeps to himself. Block two is home. He's at his desk and his mom is at the door. Same facing glasses, two outfits, the navy hoodie for the school, the gray sweatshirt for the home. His room gets the same trick as the field, a 3/4 angle
Speaker A
from the corner. In the classroom, I lock that angle on him in the prompt instead. For the classroom, I run two takes. Take two is the one. The reverse comes in low and close, and he's alone in the frame. On to block two, his home.
Speaker A
He's bored at his desk staring out the window and that's when his mom comes to the door. I run two takes and keep the second one. The window light falls warm on his face and his mom is right there
Speaker A
in the doorway behind him and I'm putting them back to back. Next, something has to hurt him. The problem is block three. One scene at his locker and it all hangs on one prop.
Speaker A
Someone's left him a note saying nerd. One word and you know how the other kids see him. Like any prop, the node has its own asset. Text is where the model slip.
Speaker A
So the word always starts from this one image. Since it's just an image, I run eight takes and keep the last one. The word nerd comes out clean and the caricature has his glasses and the game pad. The locker scene took two takes.
Speaker A
The second one's better. You can already see the note on his locker as he walks up. So, the close-up lands. Here's the finished scene.
Speaker A
I'm also skipping two of his scenes, the gym and his first gaming room. Block four is his choice again. One scene, how he answers that node. He walks back through the school in his esports jersey. Before, nerd was a label they
Speaker A
stuck on him. Now, Nerdy 23 is his gamer tag on his back and he picked it. I want the scene to start behind him so we can read Nerdy 23. Then the camera tracks in front of him as he walks and we end on
Speaker A
the other students noticing him. Two takes of this one. The walk works, but check his back. So Nerdy 23 turns into all caps. So now the jersey is yelling the prompts already specifying the name correctly. So I'm not going to rewrite
Speaker A
it. I'll just check each take against the outfit sheet. Okay, this one keeps the exact lettering. Here's the walk with that jersey. We read it, then cut to his confident face.
Speaker A
In block five, the resolution is just one scene, the tournament. It's a bigger room with stage lights and his teams in matching jerseys. I bring in the stage and one sheet with all four teammates because again, without it, Cedense kept
Speaker A
generating random people on the background. Here's one of my first takes. Without that sheet, he looks right, but the guys around him are random kids in black hoodies, not his team. I'm going to type in my request.
Speaker A
Then I take that prompt into Cinema Studio two batches and hit generate. Okay, from the first egg, I'll get the wide shot, his face, and the arena. And from the second one, I'm taking the headset, the hands, and the smile.
Speaker A
Here's the result. There's that smile. At school, he was alone. Now, he's part of a team. And that's it for the school boy. All three stories are done. So, step three is cutting them into one film. This is
Speaker A
where that four beat script from the start really helps. I can cut between the heroes when they hit the same kind of moment of the story. That order is crucial. If a story doesn't work on its own, intercutting won't save it. It just
Speaker A
hides the flaws for a few seconds. When you do mix them, every single cut needs a clear risen. Match the beats. A setup matches a setup. In the first 17 seconds, we meet all three. The ballet studio, the doctor, the classroom. A
Speaker A
problem matches a problem. The goalkeeper falls. Cut to the ball hitting the boy in the gym. A choice matches a choice. He cuts off the brace.
Speaker A
The boy puts on his jersey. She picks up the sneakers. Three decisions back to back. A resolution matches a resolution.
Speaker A
The stadium, the tournament, the battle, and at the peak, his fingers on the keyboard cut to her feet on the floor.
Speaker A
Here's one from the resolution. The ball in the goalkeeper's gloves, then the school boy's hand on the mouse at the tournament. I also used the details I generated. The knee brace, the jersey, the shoes. I cut on those close-ups so
Speaker A
the clothes and gear stay part of the action. All right, before we finish, let's watch the full ad one more time.
Speaker A
If you keep playing on this knee, you may not walk properly again. Don't tie your life to the street. I want to take care of you.
Speaker A
I'll take care of myself. A heat. And here's the most asked question. How much did it cost? This project cost about 8,500 credits counting the takes I didn't use. It's a bit pricey, but for a commercial like this, it's fine. A real
Speaker A
shoot with three casts, stadium, a dance battle would cost hundreds of thousands of dollars. So if you make a commercial like this for a brand, you can put the credit cost into the production budget.
Speaker A
The whole project took 72 hours. Unreal said this would take months from start to finish. What I take into the next one is approaching the story as a real director. So when you generate something, you already have a vision of
Speaker A
how you're going to make the story work. If you get stuck, jump into a Discord server to get help and share what you're making. This whole project started from a viewer comment. So, drop your idea for the next commercial below. If one hits
Speaker A
1,000 likes, I'll bring it to life and break down the entire process. And as always, if you found this video helpful, hit the like button, subscribe, and I'll see you guys in the next one.
Topics:Seedance 2.5AI commercialHiggsfield AICinema StudioAI video workflowcharacter consistencyClaude promptsAI filmmakingSoul CinemaCream edit
