Hey guys, before we get into it, this video is a little different. You're not getting me today. We've got someone new taking over for this breakdown.
And honestly, I think you're going to like her a lot more than me. She's been here at Cakesfield behind the scenes for a while now. And today, she's finally in front of the camera where she belongs.
So, be nice to her. Tanya, take it away. >> Five characters, one city under siege.
every single frame AR generator. This is and that's exactly how it was built. Hi, I'm Tanya and oh my god, I'm so excited to be here.
What you just saw, that's Zephr. Five characters, a monster invasion, zero traditional production, fully built inside Hicksville using Cedance 2. 0.
And today I get to show you exactly how our directors pulled it off. We're going to cover three things. How the characters were built, how the world was built, and this is the part I personally cannot stop thinking about.
How every single shot in the trailer was directed and generated. Let's get into it. Part one, building the characters.
Okay, so before anyone generates a single frame of a video, you need actors. Real, consistent, yours. And this is where most people blow it.
Here's what happens. You generate a character, you love her, you start prompting scenes, and by scene three, she looks like a completely different person. Different nose, different vibe, completely different energy.
That used to be the problem. The director solve it with a three-step workflow, and once you see it, you will never go back to winging it. Step one, sold cinema for the face.
They used one base prompt to generate all five girls. Same prom, different generations, fighting the right bone structure, the right eyes, the right presence. There she is.
That's Haramin, our youngest Arman. You can already feel that energy just from a still image, right? Step two, sell a cinema for the outfit.
Prompt it. Check it. If something disappears between the generations, like a belt or a strap or a specific accessory you want, you catch it here and add it back explicitly before it becomes a problem across photo shots.
Now, face is locked, but she needs a look. And here is something the directors figured out that I personally love. You don't generate the face and the outfit in the same prompt.
You build them separately. More control, more specificity. And here's the outfit prompt for Haru.
Notice the band-aid on the nose is explicitly in there. Such a tiny detail, but that's a detail that makes her harumin and not just a girl in a yellow top. Details anchor characters.
Generic bombs drift. Step three, fuse everything in Nana Banana Pro. Face asset plus outfit asset go in.
One production rated reference image comes out. The directors called this the screen test. In traditional film, you shoot a screen test so every department knows exactly what the actors look like on camera.
This is the AI version and this is what Cedance checks against every single time how to appears on screen. The workflow. Sol cinema for the face.
Sol cinema for the outfit. Nana banana pro to fuse them. That's your actor.
That pipeline ran five times. Five master assets. Cast is locked.
Let's build them a world. Part two, building the world. Okay, so the directors needed three things.
The abandoned city, the monster creatures, and the mechs. I learned through it fast because honestly, the animation section is where your brain is going to explode, and I want to get in there. For the city, the brief was really specific, not dark apocalypse energy, eerie warmth, golden hour haze, like the city is being slowly reclaimed with something.
The directors ran the concept through Claude first to expand the prompt, and this is something worth doing. By the way, Claude is incredible at turning a mood into actual cinematography language. Here's a piece of what came out.
First pass, well, that's close, but not destroyed enough. They push harder on debris, on overgrowth, on time having passed. Second pass, that's the CD.
The monsters went through something fun. They generated multiple creature variations, then blended the best ones together in Nana Banana Pro. And this is a tip worth saving.
If you like a generation, but it's not quite lending, combine it with another generation you made. You will get something you could never have prompted directly. The mech's edge got a color palette tied to the girl's personality.
Ra has lavender, dark blue, and silver. Dear has white and teal. When you see the Mac, you should know whose it is before the cabin opens.
That level of intentionality is what makes it feel like a real production. Okay, world is built. Cast is ready.
Let's talk about the part I've been dying to get to. Part three, directing the animation. All right, almost 100 generations went into this video.
Hundreds more didn't make it. So, how do you plan that? The director's head, one simple rule.
Every character gets introduced twice. Once in the cabin to establish who she is, double penetration. Will a third one fit?
And once in action to establish what she can do? That structure was planned before a single generation ran. First, Haro.
Her mech is down, monsters approaching, and she's lying on the cabin floor with a lollipop still in her mouth, completely unbothered. That contrast is the character. Here's the exact prompt, >> girls.
I think I need some backup. >> Notice what the prompt is doing here. It's not just describing what she looks like.
It's describing how the camera behaves. The jaws, the reestablishment, the stabilization after impact. That is the direction.
That's the difference between a cool still and an actual scene. Oh, and I love those facial expressions and the lollipop. First try, she was there.
Next, Zero. Zero's intro is all about the scale. You didn't see her first.
You see the mech. this massive white and teal machine striding through ruin city kicking out dust. The camera dragging load and wide just to capture how enormous it is.
And then it cuts inside the cabin and there she is blonde small demon horns completely calm like this is a Tuesday. And here's how the directors wrote that. >> Spotted her sending a mark.
Let's move on to Raina. The coolest one in my opinion. She looks arrogant and independent.
And that is not an accident. Those traits are literally written in the prompt because Sedance understands personality and is able to deliver the right emotions if you give it the right language. >> I hope I will be MVP today.
>> That line I hope I will be the MVP today is written into the prompt as stage direction. See dance generates delivery. And the recoil on the tail at the end, hair blown back, hands firm.
That's not describing a shot. That's describing force moving through a body. That's what gets you a real reaction instead of a character just sitting there while something explodes of a screen.
Next, Naomi. She's the funny one in the group. Honestly, I was sure the model would nail her intro, but I was still surprised by how it landed.
The emotion is right. The bit is right. The jaw lands exactly the way it should.
Sen's just read it. First try. Double penetration.
Will a third one fit. Last to show up. Mera.
Her friend is in trouble and she needs to run. This was a shot the directors really wanted to get right because a machine that size moving at speed has to feel heavy. Not bouncy, not floaty.
Heavy. >> Sorry for being late. And what sided gave back was exactly that.
The ground buckling slightly under every step. Not exploding, not bouncing. The way it actually would under something that massive.
That is physics. And that is why the directors chose sea dance for every action scene in this trailer. Now the fight scene because this is where seed dance proves itself.
Raina's cannon sequence. First, her mech is standing in the ruins. She's not panicking, fighting, and just shooting the creatures.
The directors needed this to feel overwhelming, like the monsters are everywhere, and one woman with a giant gun is deciding to be done with it. Here's the prompt. Now, the punch.
Mera saving Haru. Mira's mag runs in and punches a monster that's about to kill Haru. The punch needs impact.
And this is where so many people get it wrong. Most people would write something like me punches the monster. That's not direction, that's a caption.
The directors of Zephyr wrote the physics. Feel the difference? The weight of the body, the consequence of the impact, the specific visual aftermath.
That is what force assisted dance to generate physics instead of just movement. When you write what force does to your body, the model understands mass. It understands momentum and you get a punch that actually lands.
And then the K-pop sequence. This is everyone's favorite part. Five characters synchronized, moving as one.
This is genuinely hard for an AI video. Most models want to focus on one subject and let the others drift. But the directors had two things working in their favor.
First, the music. They generated the track separately, then uploaded it into seed dance as an element along with the lyrics. Seed dance had the bit, the rhythm, the energy of the song before it generated a single frame of movement.
The music wasn't added after. It was part of the generation. Second, they get out of the way.
Here's the prompt of that sequence. That's it. No step left, arms up, turn.
One term of actual hierography direction. K-pop. C dance knows what that means.
It knows what synchronized formation energy looks like. what the body isolation feel like, what the camera should be doing. The less they controlled, the better it performed.
And for the closing phrase, the new monster is coming. All five information. Smoke rolling in.
Lights hitting them from above. When that came, everyone stopped. That's the sequence.
That's the closing shot. And it was generated, not choreographed. So, let's sum up.
Solo Cinema for faces and outfits. Nana Banana Pro to fuse them into master assets. cloth to build out the world and Cedense 2.
0 to direct everything. Personality shots, fight scenes, dialogues, choreography. And remember, this is just the intro to Zephyr.
There is lots more to come, so make sure to stay tuned. And as always, if you found this video helpful, hit that like button, subscribe, and I'll see you guys in the next one.
