Quick experiment 01 · Fashion film
Knight in Bloom
The soft side of a knight. An editorial still, a 30 second fashion film, a living poster, and a test of five video models.
4 hrs
start to finish
3
rounds to the hero still
$48
in generation, all in (USD, about)
5
video models tested on one b-roll prompt
Overview
An editorial still of a knight asleep in a flower field, and a 30 second fashion film that shows how he got there. I also made an 8 second living poster loop and ran a b-roll test across five video models. It's the quickest thing I've made.
This was a quick experiment, about 4 hours from start to finish. It isn't meant to be perfect, and you'll spot a few AI flaws along the way. I wanted to see how far I could get in a short time. I directed. Claude acted as producer and ran the video generation through my Leonardo API. Most of the costs below come from the API log. The rest are my estimates.
01 · The brief
Masculine and soft at the same time
I wanted it to feel like a high end cologne ad. Masculine and soft at the same time, two meanings in one world.
The idea it hangs on is curiosity. Why is he in armour? Why does he look so peaceful? What's the story with the flowers? The film answers those questions.
02 · The numbers
What it took
Time
About 4 hours. The hero still took about half an hour of that
Generation cost
About $48 all in. From the log: about $2 for 15 keyframes, $24 for the film, about $8 for music, about $6 for the loop, $1.44 for the Flux 3 b-roll. Estimated: about $2 to $4 for the other b-roll models, about $1 to $3 for the stills I made in the Leonardo app
Stills
25 kept, from about 35 generated. Hero: 3 rounds. Film: 15 keyframes, 8 polished frames, 1 character sheet
Clips
Film: 1 run, used. Loop: 1 run, used. B-roll: 5 models, 1 clip per model
Music
About 10 cues across cello, baroque, suspense and violin moods, 1 used
Models
GPT Image, Seedream 5.0, Seedance 2.5, Leonardo music. B-roll test: Seedance 2.5, Flux 3, Wan 3.0, Gemini, Hailuo. Edited in Premiere Pro
Logged figures are what the Leonardo API charged for the runs Claude made. The app stills and the other b-roll models are estimates. None of it includes my own time.
03 · How it was made
Five steps, from a loose idea to a model test
1
Look first, subject loose
My vision was loose on purpose: a male model, blonde buzz cut, armour, a field. So I worked backwards. I found dreamy film photography references, uploaded them and asked for that look with a male model. The first image was him, so I didn't explore further.
GPT made it too warm, so I took it into Seedream 5.0 to cool it down, zoom out to show more flowers, and make it square. Closed eyes and leaf shadows on his face also deal with the two things that give AI faces away: dead eyes and plastic skin.
2
Character sheet
Front, back and close up from the hero image. Very low effort.
The character sheet, made straight from the hero image.
3
Story into keyframes
I plugged my Leonardo API into Claude and gave it the story: he walks into the field, takes his armour off piece by piece, lies down and relaxes. Claude wrote the shots and made the keyframes.
Then the fixes. He came out in a long skirt, then a skirt again when I said tight, so I said pants. They're a bit theatrical, and for a perfume ad I'm fine with that. The dropped armour looked like a shop display, so I asked for it scattered. And my final photo shows him still wearing his breastplate, so the film could only drop the gauntlets, sword and helmet.
The gauntlet kept coming out with three fingers. Editing it failed twice and made the image worse, so we started fresh and described how a gauntlet is built: four finger plates and a thumb plate. That fixed it.
Before · long skirt, armour laid out neatly
After · pants, armour dropped as he walks
Wardrobe and the armour trail, before and after.
Before
After · four finger plates and a thumb plate
The gauntlet, fixed by describing how it's built.
4
The film and the sound
One Seedance run gave me what I needed. The sound sheds metal too: every armour sound is quieter than the last, until there's only wind and a breath.
For music I prompted sad baroque violin, got it on about the fifth try, and edited it in Premiere Pro.
5
Loop and b-roll test
The loop animates the final photo: leaf shadows drift, flowers move, his chest rises once.
Then I took one short wildflower prompt and ran it through five models. Flux 3 cost $1.44 and looked real. Wan 3.0 came out too saturated for this look. Gemini was almost half the price and nearly as good. Hailuo didn't hold up. If you're building a library of a hundred clips, those price gaps add up fast.
The living poster, 8 seconds, looping.
Flux 3 · $1.44. Looked real.
Gemini · almost half the price, nearly as good.
Wan 3.0 · too saturated for this look.
Hailuo · didn't hold up.
One wildflower prompt, four of the five models.
04 · Looking back
What worked, and what I'd change
What worked
Starting loose worked. I only set a few rules: a male model, blonde buzz cut, armour, a field. The AI was free to invent him, so the hero still took three rounds.
What I'd change
What I'd change is the gauntlet. I'd describe how it's built from the first prompt and skip the edits that wore the image down.
Frames from the film, flower to flower.
Credits
Made with
Images and video made with the Leonardo API and the Leonardo app. Claude as producer. Edited in Premiere Pro.
Work with me →
© 2026 Phyo Thu · @phyofox
Wellington, NZ