Home / ModelScope

ModelScope

2026 · daily generative video · 1,672 clips so far

ModelScope clips come from the other half of the system. They are generated from my own tracked data rather than the news: the same daily record Surveil Thyself turns into posters, pushed through the language and then through a 2023 text-to-video model.

How a day becomes a clip

The day is collectedAt 07:45 every morning the collector writes down the previous day from the traces left behind during ordinary use: photographs, pages browsed, videos watched, notes kept.
The language reads itIt compiles all of that into a dossier, the same kind of profile an ad company builds on you, then hands the dossier to z0glyphic 913 and to a video-generation model.
Prompts are written from the readingThe reading becomes a run of short scene prompts. Every prompt ends with the same look: VHS tape footage, heavy analog grain, low resolution, cross-processed color, distorted motion, lo-fi documentary, surveillance camera perspective, 1990s.
The model makes the clipsModelScope (damo-vilab, text-to-video, 1.7 billion parameters) runs on my own computer. Each clip is three seconds at 256 pixels; one takes several minutes.
They are cut togetherThe day's clips are assembled into one 24-second piece with a generated soundtrack.

The model was trained partly on stock footage, so watermarks and half-remembered stock gestures surface in the results. I leave them in.

It sits next to the Oracle as the other daily automated output: one turned outward at the news, the other turned inward at my own habits. Nothing about a clip is chosen by me.

The prompts themselves are not shown here, because they are written from my private record.