Intelligence
begins with
seeing.
A research institute investigating visual intelligence beyond the limits of language.
INsVI brings fine arts, machine perception, spatial reasoning, and data science into one field of inquiry.
What we investigate
Visual
intelligence
How machines form spatial, compositional, and perceptual understanding—not merely captions about images.
Visual data
as culture
How datasets encode artistic practice, cultural judgment, absence, bias, and ways of seeing.
Art after
generative AI
How artists can test, critique, and reshape intelligent systems through visual experimentation.
Images are not
sentences in disguise.
Language models describe the visible world. Visual intelligence must learn its structure, relations, ambiguities, and force.
Visual Intelligence Beyond the LLM
Large language models can name, caption, and discuss images, but that is not the same as seeing as an artist or designer sees. INsVI studies visual intelligence as a distinct mode of thought exercised through making, selecting, comparing, composing, revising, and judging visual form.
This is not simply spatial intelligence. Spatial intelligence concerns location, orientation, geometry, navigation, and objects in space. Artistic and design intelligence also works through proportion, rhythm, contrast, hierarchy, color interaction, material behavior, ambiguity, metaphor, cultural context, affect, and aesthetic judgment.
Read the current inquiry
Beyond language-mediated vision
Multimodal systems often privilege what can be verbalized. They may recognize what is depicted while failing to understand why a composition works, which subtle alteration weakens it, or how an artist discovers meaning by looking and making before that thought can be stated in words.
LLMs and visual evidence
Fluent language can conceal weak visual understanding. Captions, labels, embeddings, and preference scores can flatten texture, scale, sequence, context, and relations among parts. For AGI, the question is not merely whether a system can talk about the visible world, but whether it can form, test, and revise visual judgments without treating language as the sole measure of understanding.
The dataset is part of the intelligence
Web-scale datasets contain duplication, poor reproductions, missing provenance, inaccurate captions, cultural imbalance, uncertain rights, and images detached from their material setting. A serious art-and-design dataset should preserve authorship, provenance, medium, dimensions, viewing conditions, process, revision, relationships among works, interpretation, and disagreement. Dataset construction is a curatorial and epistemological practice—not merely an engineering task.
What INsVI is seeking
We seek to articulate the visual intelligence used by artists and designers: the capacity to discover meaning through form, organize attention, perceive relationships before they become propositions, and make judgments whose evidence remains substantially visual.
Institute for Visual Intelligence
Visual intelligence research · New York