What “code-generated video” means for launch films

“Code-generated video” is a new name for something that has existed for years: a film drawn by a program from a plan, instead of painted by a model. The name matters now because more tools make it, and because it answers a question every founder asks before a launch: can I trust what the video shows? This post explains the term, what changes for a launch film, and where it stops.

Two ways to make a video from words

There are two quite different ways to turn a description into a video. The difference is what the machine is asked to produce.

• Pixel generation. A diffusion model starts from noise and paints each frame of a clip. It is good at footage, people and places, and it gives a different clip every time you ask.
• Code generation. A plan says what is in each shot (type, shapes, screens, a camera) and how each piece moves over time. A renderer draws the frames from the plan. Ask twice and you get the same frames.

A language model can write the plan in both worlds. What differs is who draws the picture: a diffusion model, or code.

What changes for a launch film

A launch film is mostly words, interface and one or two numbers, watched by someone scrolling with the sound off. That is the work a drawn film does well and a painted one does badly.

• The text is exact. A word is set as real type, so your product name is spelled the way you wrote it, at any size.
• The screens are yours. The interface in the film is your real screenshot. A drawn film has no reason to invent a button.
• Edits are small. The tagline moves, a screenshot is replaced, the length is trimmed. Each is a change to one shot, and only that shot is drawn again.
• The result is repeatable. The same plan gives the same frames, so a film you approved does not drift when you render it again.

How LaunchReel does it

LaunchReel has made one kind of film since the beginning: the product launch film. The pipeline has six stages, and the first two are plain text you can read before a frame is drawn.

1. Brief. It reads your site and answers six questions about the product, each with its source page, then writes the brief.
2. Storyboard. Every shot gets a duration, its on-screen words and the motion it will use.
3. Components. Each shot is built from named motion components: transitions, titles, device frames, diagrams, typing effects.
4. Music grid. The film carries a beat grid from the tempo of its music bed, and components that can land on a beat are placed on it.
5. Narration. One or two lines per shot, in English or Chinese, each voiced with the neighbouring lines as context.
6. Render. A headless browser draws the frames and ffmpeg encodes the MP4, with captions burned in.

The full walk-through, with the film samples, is on the page about code-generated video. The motion layer is open: the components are on compreel.com/motion and the package is MIT licensed as @launchreel/motion-kit.

Where it stops

Drawing from code has limits, and they are the mirror of its strengths. It is not for live-action footage, photoreal scenes or 3D product renders. It is made for software and web products. A film is about 30 seconds, and narration is in English or Chinese. If your launch needs a film of people in a room, you need a camera, not this.

How to judge any tool in this lane

Whoever makes it, five questions tell you whether a tool really draws from a plan:

• Is the text real type, and does it stay exact at small sizes?
• Does it show your real screens, and does it say so when it does not know something?
• Can you change one shot without losing the rest?
• Are the effects named, so you can ask for a specific move?
• Can the result leave the tool in a form your editor can use?

If you want to see the method on a real site, the use-case pages show how it reads for a SaaS, an app, a shop or a developer tool, and the product launch video template lays out the shots second by second.