A world begins with a few words
A little world, in your words.
A quiet walk through the woods.
This is a previously generated example. Your new video will appear here when it is ready.
Record a camera path, then generate. New frames take minutes on this Mac Mini; the example is playback.
See timings ↗Behind the world
Back to studio ↑The film / Hy-Nomas on Apple silicon
A world, continued.
Edited product film with real generated footage. Generation is offline; camera controls were tested as pre-set inputs.
Download the trailerAbout the footage and measurements
The film combines actual Hy-Nomas generations with motion graphics showing text conditioning, video latents, decoding, and fused Metal operations. A matching three-second generation took 19 minutes 51.41 seconds on the M4 Mac Mini and 4 minutes 17.39 seconds on the M3 Ultra Mac Studio: 4.63 times faster. Both used cached text, the same prompt and seed, 72 frames at 1280 by 704, 24 fps, and eight steps. All video latents and 18 saved pre-encoding frames matched exactly. A separate six-second W-to-D camera and prompt continuation test took 10 minutes 41.50 seconds. The scene remained coherent, but the requested stop-and-look action was not followed. No live keyboard response is demonstrated.
Music: “Mechanolith” by Kevin MacLeod · CC BY 4.0. Excerpted, faded and mixed.
Run it yourself
Your Mac. Your scene.
Get the optimized MLX runner and a step-by-step guide. Setup downloads the pinned model weights from Hugging Face; inference runs on your Mac.
Measured for a 3-second clip · 1280 × 704 · 24 fps · 8 steps · same prompt and seed · cached text.