AI VIDEOS

The clip you code.

Moonshot's Kimi K3 arrived this week with a three.js world as its calling card. A working note on letting a coding model build the scene, so the camera, the light, and the take stay yours.

TLDR

A new open-weight coding model writes believable three.js scenes on the first pass. That reopens code as a second road to moving images, beside the video models rather than instead of them.

•  Moonshot AI released Kimi K3 on July 16, an open-weight model with a million-token context and native image understanding, usable on its site and through developer access now.

•  Its launch demos included a browser 3D open world written in three.js with WebGPU, with forests, a village, snowy mountains, and moving weather.

•  For creators the point is practical. Describe a shot in plain words, let the model write the scene file, run it in a browser, and screen-record the take.

•  The caveat: it is a coding model, not a video generator. The look is geometric rather than photoreal, the full weights are promised by July 27 but not out as this is written, and the specs are the vendor's own.

•••

There is a second road to a moving image, and it reopened this week. Moonshot's Kimi K3 landed on July 16 with the usual chart talk, and one demo that said more than the charts: a small 3D world, running in a browser, that the model had written as code. Forests, a village, snow on the ridgeline, weather drifting through on a loop.

A video model hands you footage. A coding model hands you a scene: geometry, a camera, lights, and an animation loop in one three.js file. The difference is not quality. It is control over time. Code plays back the same way every run, and a take that repeats exactly is a take you can direct.

I have been treating that demo as a craft opening rather than a novelty. Not every clip wants to be footage. Some clips want to be built.

WHAT LANDED

A coding model with a camera.

K3 is Moonshot's new flagship, open weight, with a context window that runs to a million tokens and native image understanding. It is reported to sit near the top of the field for long-horizon coding, the kind that holds a plan across a whole file rather than a snippet. Standings move weekly, so hold that loosely. What is not in dispute is the demo: a working open world in three.js and WebGPU, produced by the model.

You can use it on the Kimi site today or through developer access, and the full weights are promised by July 27. For this issue none of that matters as much as the shape of the demo. The proof was a place, not a paragraph.

THE IDEA

Direct the scene, not the pixels.

The move is to brief the model the way you would brief a camera operator, not the way you prompt a diffusion model. Name the subject, the camera move, the duration, and the light, then ask for one file with one animation loop. The vaguer the brief, the more the model decorates. The tighter the brief, the more the scene looks decided.

A low-poly island at dusk. One slow orbit, twenty seconds. Warm key light from the west, thin fog over the water. One three.js file, one animation loop, nothing beyond three.js.

Paste the result into an HTML file, open it in a browser, and the shot plays. No render queue, no seed lottery. If it is not right, the next section is the whole job.

THE REVISION

Change one named thing.

Here the coded clip pays for itself. In a video model, a revision is a new roll of the dice and the whole frame can drift. In code, a revision is a line. Ask for one parameter by name and everything else holds still. Fog density, orbit speed, palette, the hour of the light.

Keep the scene exactly as it is. Halve the orbit speed. Raise the fog one step. Nothing else moves.

Run a few passes like that and you are not generating anymore. You are arranging a shot, the way an editor nudges a cut. Keep each pass to one change so a broken pass is easy to blame and easy to undo.

THE EXPORT

From browser to deliverable.

The delivery path is plain. Run the scene full-screen and screen-record at your target frame rate, or ask the model to add a frame counter and capture stills for compositing. Because the loop is deterministic, a retake costs nothing and matches the last one exactly, which is something no diffusion tool will promise you.

Keep the scene file in the project folder next to your prompts. It is the same discipline as saving a prompt, except this artifact reruns forever.

THE LIMITS

Where the coded clip ends.

Honesty about the lane: three.js gives you a geometric, procedural look. Title cards, loops, worldscapes, abstract motion, product turntables. It will not hand you a photoreal actor, and chasing one here wastes the medium. Photoreal stays with the video models.

And a coding model will confidently write a scene that does not run. Small briefs, small passes, and a quick look at the browser console keep a bad pass small. The craft transfers anyway: the same director's brief that makes a clean coded scene makes a cleaner video prompt.

THE TAKEAWAY

Two engines, one storyboard.

Nothing here replaces your video model. It adds a second engine with opposite virtues: the diffusion tool is lush and unrepeatable, the coded scene is spare and exact. The storyboard decides which engine each shot deserves.

So take one shot from your current project, the most geometric one, and build it instead of generating it. Twenty seconds, one orbit, one light. You will come back understanding time in your clips a little differently.

•••

I am putting together a scene script kit: eight shot briefs written for coding models, each with the plain-words ask, the three.js scene it produces, a one-parameter revision ladder, and a screen-record checklist, with a worked example for each.

Want it when it ships? Reply with send me the scene script kit and I will get it to you.

A QUESTION FOR YOU

Which shot would you build in code first?

Reply with the scene, the camera move, and the length. The kit's briefs will start from the shots people actually want, not the ones that demo well.

If this was useful, forward it to a creator who storyboards in keyframes.

Until next time,

Luxe Prompting

Luxe Prompting

AI Image Generation for Creators