AI Aesthetics Map

A visual classification and analysis tool to map dominant patterns in video generated with artificial intelligence

The project is born from a central question that has crossed Artefacto’s research since its origin: does artificial intelligence have its own aesthetic? The «AI Aesthetics Map» tool answers this question at scale, processing a broad corpus of Instagram videos produced with generative technologies. The work is directly anchored in the findings of the academic article Implications of generative artificial intelligence in cinematic narrative and aesthetics (Caballero, Sora-Domenjó, and Codina), where the analysis of 295 cinematographic projects revealed recurring aesthetic patterns that this prototype systematizes and visualizes.

The Instagram environment, with its mix of jokes, surrealist content, and formal production videos, offers a spontaneous cultural laboratory that the academic study did not cover: the massive, dispersed, and popular production of images generated with AI.

The tool uses Google’s Gemini model through its official SDK. Each video is sent to the model along with a structured taxonomy of 25 main categories and 173 subcategories, developed specifically for this project. The model receives the video as multimodal input and returns a structured JSON object. The taxonomy operates as a classification ontology designed for LLMs, with explicit instructions on when to prioritize content over technique and how to treat videos where the generating tool (Sora, Runway, Deforum, etc.) is only a medium and not a theme.

The system connects with Google Drive for corpus management and allows direct import from Instagram.

The article by Caballero, Sora-Domenjó, and Codina identifies that most analyzed cinematographic projects with AI resort to original image generation through diffusion models (Stable Diffusion, Runway…), text-to-speech, and video scaling and enhancement techniques. But beyond the technologies, the authors point to the emergence of a series of visual features that repeat regardless of the author, country, or artistic intention. Continuous morphisms appear where bodies transform without discrete transitions, textures oscillate between hypersaturation and controlled blur, visual anachronisms fuse historical eras, diffuse backgrounds contrast with hyper-stylized central subjects, and there is a systematic presence of oneiric surrealism as the default mode of representation. These patterns raise a cultural question: does the model learn aesthetics from training data and replicate them, or is a genuinely new visual sensitivity emerging?

The prototype functions as an instrument of aesthetic cartography. From the corpus, it produces distribution maps that allow identifying which categories predominate, which aesthetic subcategories accumulate in certain thematic zones, and what relationships exist between the type of content (humor, conceptual art, technology, nature…) and the visual features generated by AI. The tool does not intend to evaluate the quality of the videos but to reveal their collective visual grammar. Within Artefacto’s research, this map functions as empirical evidence for the hypothesis posed in the academic study: that generative AI is producing an identifiable, exportable, and culturally conditioned aesthetic that circulates in cinema between auteur festivals and Instagram memes with a visual coherence that neither space is conscious of sharing.

Application (requires authentication): https://artefacto-mapa-de-v-deos-hechos-con-ia-231178493219.us-west1.run.app/