Quick experiment 03 · Short film
Empty
A girl racer runs out of fuel. A dusk photo, and the 30 second short film that ends on it.
8 hrs
start to finish
$106
in API generation, including $26 of mistakes (USD)
6
film runs, 1 used
11
shots in the final cut
Overview
A Cinematic Dusk photo and the 30 second short film that ends on it. It plays like a cool street racing film. Then she drift parks, slams the wheel, and we find out she's just out of fuel. She texts her dad to come get her.
This was a quick experiment, about 8 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. Costs below only cover generations Claude ran through the API.
01 · The brief
She still has to call her dad
Cinematic Dusk had to feel cinematic because of the story, with the lighting coming second. I wanted nostalgia from Japanese car modification culture, so the car is an AE86. Right hand drive, on the left side of the road, like it's in Japan.
The idea it hangs on is the punchline: she looks like a total badass, but when she gets into trouble she still has to call her dad.
02 · The numbers
What it took
Time
About 8 hours, mostly 2 to 3 Sept
Generation cost
About $106. That includes one failed run and about $26 of test jobs that fired by mistake while setting up the API
Stills
Photo: about 8 rounds. Film: 9 shots designed as stills, plus extra rig mount and drone frames for the last version
Clips
6 film runs (1 Kling test, 5 Seedance), 1 used
Models
GPT Image 2, Nano Banana, Leonardo upscalers, Seedance 2.5, Kling 2.5 for one test. Photoshop for the photo finish
Generation cost is what the Leonardo API charged for the runs Claude made. It doesn't include my own time or the stills I made in the Leonardo app.
03 · How it was made
Five steps, from a character sheet to a text message
1
Character and car first
I built a character called Tessa from a Pinterest image, plus a car sheet with four angles and the pop up headlights. The first face came from GPT and looked oily and blotchy, the way AI skin does at low res. I generated more faces, combined them, and rebuilt the character sheet with Nano Banana.
For the interior I found a photo of the real AE86 cockpit and changed the lighting and the GPS.
First face, from GPT
Rebuilt character sheet
Tessa, before and after the rebuild.
The car: front and three quarter studio sheets, the pop up headlights, and the right hand drive interior.
2
The photo
The first version looked like a random phone snap. So I posed her sitting on the ground between the taillights, texting. An old car with a modern phone makes you ask questions.
The agent anchored the wrong girl and used the wrong base image a couple of times, and I caught both. I did the final crop, grade and grain in Photoshop.
3
Story into shots
I told Claude the story in plain English and asked for six to ten shots across 30 seconds. I took that shot list into the Leonardo agent and made the storyboard frames myself.
4
The film
Claude ran the whole 30 seconds, picture and sound, in one Seedance pass. The first version had no driver in the car during the drift. We fixed it with a new reference of her at the wheel.
Then I pushed for more energy: a camera rigged to the car, a top down drone shot after the GPS warning, quicker cuts. It went to 11 shots.
The drone keyframe, added in the last version for energy.
5
The text message
I designed the message bubble myself. Claude added it with normal video tools and a small pop animation. Graphics like this need to be pixel perfect, and AI isn't the right tool for that.
The last frame, with the message added in post.
04 · Looking back
What worked, and what I'd change
What worked
The order of the story worked. The reveal comes after the frustration, so it lands as a joke.
What I'd change
What I'd change is the references. They weren't sharp enough when we started, and the film came out soft because of it. Now I finish every reference to final quality before it goes near a video.
Frames from the film, city to lookout.
Credits
Made with
Images and video made with the Leonardo API and the Leonardo app. Claude as producer. Finished in Photoshop.
Work with me →
© 2026 Phyo Thu · @phyofox
Wellington, NZ