I made the vertical drama about vertical drama

Vertical What? is a vertical drama that explains vertical drama to people who have never watched one. Episode 1 is called Twist. An episode runs about a minute and a half, it is shot upright for a phone, and it breaks off right where you need the next one.

This autumn I updated my vertical drama booklet for MIPCOM, the big television market held in Cannes every October. I spent my days explaining the format: what a twist is, where the stories come from, who pays. At some point the obvious question arrived. Why not make a vertical drama that explains vertical drama?

So I did. Nine episodes of about a minute and a half each, on YouTube, covering what the format is, where it comes from, how it is paced, what AI changes, and who the money ends up paying. Here is how it got made, and what it cost.

The one-eyed cat

On March 3 I ended an article about spending $120 on 90 seconds of AI cats with this sentence: "I will probably do it again. Maybe that is the real problem." The hero was the one-eyed ginger cat of my novel. The video model refused to keep the eye damaged: every shot began with the injury and healed it halfway through. At about $80 a minute, I had bought a very expensive lesson in humility. I drew one conclusion, that handing a director's tools to someone who is not a director gives you a race car stuck in first gear. I did not stop.

One tool per task

I started like everybody does. An assistant, Claude Code, writes the code and runs the operations for me. An image generator, gpt-image-2, draws the sets and the characters. Kling animates the images. Runway syncs my own voice takes to the lips of a drawing. ElevenLabs does the voices, including a clone of mine.

Episode 1 cost me less than $10, about $7 a minute by a looser count. I never made a second one that way. I had asked for an image with two paws resting and one paw acting, and it came back with three paws. I counted them once, then twice, then fetched my neighbour, who counted three as well and asked whether it was an exercise. I said yes. The animation model inherited the flaw and walked the cat on three paws for five full seconds.

DramaMagic, kept without the tool

At the end of September I tested DramaMagic, a tool for making vertical dramas from BytePlus, the business arm of ByteDance, the company behind TikTok. It did not work well for me. It was slow, I could not work out how to make episodes, which means a season, and the French sounded wrong. The lip-sync in French is what made me leave.

The breakdown, though, was excellent. A script becomes shots, every shot gets its own prompt (the instruction you give the model), every prop gets its own sheet. I kept that and rebuilt it at home, without the tool. Fal, the platform that gives me access to dozens of models, has its own library for storing characters, props and sets. Mine holds 4 audio files and no video.

Seedance 2.5, and the price of a minute

For movement I moved to Seedance 2.5, a ByteDance video model that I call through fal. For my series, with characters drawn in flat colour, short shots and many close-ups of objects, it is very good. It has company: version 2.0, released on February 12, joined Kling and Vidu in the production of fully AI-made microdramas in China, according to Yahoo Finance.

I render at 480p. At fal's public rate, a second costs about $0.22 at 480p against $0.47 at 720p, more than double. Those figures use the public price list, not what I paid out of pocket, because Y Combinator gives me credits with some providers.

Episodes 3 to 6 are the only ones where I measured everything. A minute cost me $26, then $18, $18 and $14. A third of everything I generated on those four episodes went in the bin. Why did the price fall? I suspect two causes: I was learning, and flat illustrated characters are easier to render than realistic ones. I never measured how much each one weighs.

There were mistakes. A werewolf has to look like a real wolf on screen, or the model hands you a man with wolf ears. A set has to stay frozen, or an office appears while the hero is talking. And a model left to generate sound without a reference voice starts speaking Mandarin. The automated check on the images does not catch that. I heard it, and I had to tell the agent, who swore the clip was clean.

Kintu-bila, the realistic one

In parallel I had another project: adapting my short story 0-3-1-4-1-5 into a vertical drama, set in the universe of Kinza. A technician, his daughter, a drone and an unknown word, "Kintu-bila". I wanted realistic images this time, not flat colour.

The 30-second teaser cost about $75 a minute, against $14 to $26 for the four episodes where I counted every render. Realism taught me other things. Seedance refuses photoreal character images as references, on the grounds that they could look like real people. ElevenLabs refuses to create the voice of a 7-year-old girl, for very good reasons, so the girl's voice became a young woman's. The drop in price, then, holds for one style of image and not for the other.

The ADSL

My connection is ADSL, 1 Mb/s upload on paper. On September 25 I measured it for real: 4.4 KB/s. I cannot explain the gap. At that rate a 135 MB film takes close to 9 hours to go up, and my first four episodes together would have taken more than a day.

The fibre was supposed to arrive a month ago. The first technician came and said no. The second said I needed a cherry picker. The third said nobody had warned him about the cherry picker. I now have a fibre installation in progress and a cherry picker somewhere in the wild.

So I did what someone who can no longer send anything does: I stopped sending. I rented a $7-a-month server, installed my assistant on it, and delegated the renders, the storage and the publishing. The films never pass through my Mac. I approve from my phone, over 5G when needed. I did not decide to work with agents from day one; the line decided for me.

The process, and FFmpeg

My tool list changed five or six times in two weeks. A tool that looked indispensable on Monday was replaced by Thursday. What survived is four habits:

• I write the script in shots, not scenes: one shot, one image, one sentence.
• Every prop that comes back gets its own reference image.
• I render one shot at a time, with a ceiling announced before I pay, and I look at the result before buying the next one.
• I edit with free tools, and an agent publishes with my green light.

That agent is called Zernio. You give it a film, a title and some accounts, and it pushes the film to YouTube and TikTok. It is free for two social accounts, so I am running on the cheapest plan.

One thing never changes: FFmpeg, the free software that can do almost anything to a video file. Cut, paste, mix, handle the sound, set text. I have used it for 25 years, and I even made a game of it, ffmpuzzle, a Wordle where you guess the settings of an encode. It also spares me DaVinci Resolve Studio, the professional editing software, which costs close to $300. Making vertical drama means redoing FFmpeg every day, and I enjoy it a lot.

Am I an artist?

In March I wrote that a real artist would have got ten times more out of the same tools. Nine episodes later I know how to make something, and I still do not know whether I am a director. In between I won a small literary prize, the Alain Le Bussy prize, in an open call for texts, for my short story "Lume Saudade", published in issue 96 of Galaxies. It does not really answer the question.

The full making-of, in French, is on Streaming Radar.