A Large-Scale Computer-Vision Mapping of the Geometric Structures of Stroboscopically-Induced Visual Hallucinations
Ethan Grove, Trevor Hewitt, Anil K. Seth, Fiona Macpherson, David J. Schwartzman
bioRxiv (Cold Spring Harbor Laboratory) February 18, 2026 preprint DOI: 10.64898/2026.02.18.705710 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Observational study |
|---|---|
| Population | Attendees of Dreamachine, a public installation |
| Intervention | stroboscopic light stimulation |
| Keywords | Visual hallucination Phenomenology philosophy Pipeline software Visualization Pattern recognition psychology Visual perception Artificial neural network Artificial intelligence Computer vision Object grammar Cluster analysis Simple philosophy Dimensionality reduction Deep neural networks Feature linguistics Geometric shape Locality Cognitive psychology |
| Citations | 1 |
| Key findings | An unsupervised computer-vision pipeline identified recurrent visual motifs in drawings made after stroboscopic light stimulation, including both known and novel geometric forms. |
Abstract
Abstract Visual hallucinations (VHs) occur across psychedelic states and diverse psychiatric and neurological conditions, yet their phenomenology remains difficult to characterise. Empirical research on VHs is hindered by the lack of large-scale phenomenological datasets, which limits both mechanistic accounts and the systematic characterisation of when and how they arise. Stroboscopic light stimulation (SLS) viewed with closed eyes provides a reliable, non-pharmacological method of inducing VHs in healthy populations. These hallucinations typically consist of vivid colours and dynamic geometric patterns that resemble simple VHs described in both psychedelic and clinical contexts, suggesting partially overlapping neural mechanisms. We developed and applied an unsupervised computer-vision pipeline to analyse a large dataset of 10,598 drawings made following exposure to hallucination-inducing SLS. These drawings were produced by attendees of Dreamachine, a large-scale public installation designed to elicit stroboscopically induced visual hallucinations (SIVHs). We extracted feature embeddings with a self-supervised deep vision transformer, then applied dimensionality reduction and density-based clustering to identify recurrent visual motifs in a data-driven manner. The majority of drawings contained geometric forms, consistent with prior observations of simple VHs under SLS. However, we also identified novel and underreported geometric formations, such as concentric squares, crosses, hyperbolic patterns, and other geometries. Our results show how an unsupervised computer-vision pipeline can organise large, openly shared phenomenological datasets into interpretable classes. By mapping the diversity of simple geometric VHs at scale, this work places new constraints on existing theoretical accounts and motivates targeted experimental work linking SLS parameters, neural dynamics, and geometric visual hallucinations.