A Python script reads the file yolov5x6.pt as if it were text, even though it was never meant to be read that way. Most of the file does not resolve into readable characters, so what remains is a stream of garbled symbols broken by the occasional fragment that survives. This is shown as scrolling text rather than run as the model it is. The file is read, but never used for what it was built to do.
The text scrolls upward in plain grey on black. There is no colour in the file itself. As each frame appears, the script runs a set of old detectors across it, looking for frontal faces, profile faces, eyes, smiles, and upper and lower bodies. These detectors come from a method developed in 2001, the standard way of finding faces before the current generation of systems replaced it, and they are still included in the open source vision library most people reach for today. They are living fossils, an older form of machine vision still in everyday circulation. Pointed here at scrolling text made from a model’s raw data, every detection is wrong. Each one is drawn on screen as a labelled box. The smile detector fires hardest. The screen fills with blue boxes marking shapes it reads as smiles, so many that they overlap and bury their own labels and one another’s. The original text is all but hidden. Now and then a face or body box surfaces and holds long enough to be read before it too is lost in the field of smiles.
These false detections are a machine version of pareidolia, the human habit of seeing faces in clouds or wood grain. The mechanism is different, a detector returning a match rather than a mind projecting one, but the result is the same shape: a face found where there is none. The boxes record a likeness that was never there, only projected. What these detectors are built to find is human likeness, and pointed at noise they go on looking for it in data that holds no trace of it. The result is a hallucination passed between machines: one generation of vision models reads the raw material of the next and finds only people who are not there. Neither the data nor the detector holds a single human figure, yet the screen fills with smiles, a program grinning back at a file that contains no one.
The detections are the only source of sound. Nothing in the data is turned into sound directly; every sound comes from a false detection. Where a detection sits across the screen sets its pitch, from low on the left to high on the right, and where it sits in the stereo image. Each kind of detection has its own voice: faces produce sustained, held tones, smiles a metallic bell-like sound, whole bodies a deep bass, and upper and lower bodies two kinds of percussion. Because the smile detector fires constantly, those bells layer and overlap into a dense, ringing mass, the sonic equal of the blue boxes covering the screen. Eye detections are the opposite case, held back by a filter so they are rare and almost silent. Every sound fades in briefly to avoid a click at its start. An echo bounces each sound between the left and right speakers, so detections gather into a slowly shifting field. The loudest passages are smoothed rather than cut off sharply.
YOLOv5 is a system built to find and name objects in photographs in real time. Where an image generator such as Ideogram 4 builds pictures out of noise, YOLOv5 was trained to locate things already in front of it. The two kinds of system leave different traces in their files, but neither trace is readable in what this script produces. Unlike Ideogram, YOLOv5’s training data is public and documented. The way the file is read here has nothing to do with what the model would normally do. It is a deliberate misreading, with its own logic, that ignores the rules of the file format and of the model it comes from.
YOLOv5 is open source, made by a company called Ultralytics. The model, the data it was trained on, and the method used to train it are all public. There is no interface to get around. Reading the file as image and sound rather than running it is not a matter of access but of choice.
This excerpt is one minute taken from the opening of the 269 MB file.
The video is 1080 × 1080, at 30 frames per second.
#unindexed #nearinference #openweights #softwarecinema
