Epistemic Cultural Flattening in Generative Visual AI: Benchmarking Hungarian Heritage and Designing a V4 Path Toward Culturally Aware Text-to-Video

Abstract

Generative image systems increasingly shape how culture becomes visible in design workflows and heritage interpretation. Their outputs often achieve technical plausibility while offering limited support for validating cultural provenance, shaping how synthetic images circulate as cultural references. This article introduces Epistemic Cultural Flattening (ECF) and an Epistemic Interpretive Framework (EIF) to distinguish structural performance from epistemic readability and to describe reductions of culture-specific legibility under globally dominant visual templates. The study operationalizes EIF through a cultural fidelity benchmark rating generated images by cultural fit, stylistic accuracy, and technical quality. It uses a Hungarian heritage benchmark set within a cross-cultural comparative corpus and compares outputs from four diffusion-based generators. The article proposes an ECF failure-mode typology that makes cultural flattening visually legible. It also outlines a V4-oriented workflow for culturally aware text-to-video, integrating GLAM sourcing, multilingual metadata, controlled model adaptation, and expert review for low-resource cultures in Central Europe.