Visions Of Destruction: Exploring Human Impact on Nature by Navigating the Latent Space of a Diffusion Model via Gaze
Mar Canet Sola, Varvara Guljajeva
TL;DR
Visions of Destruction investigates how gaze-based audience interaction can reveal human impact on nature within a diffusion-model artwork. The authors implement a real-time latent-space navigation pipeline where an eye-tracker defines a mask that guides inpainting with Stable Diffusion, animated by a randomly chosen pre-defined prompt to morph landscapes from pristine to ecologically damaged states; if gaze is absent, the scene regenerates to pristine beauty. The paper places the work in the lineage of eye-tracking and neural-art avant-garde, arguing for embodied interaction and meaningful human control over AI processes. It offers a concrete, reproducible framework for gaze-driven, latent-space manipulation that foregrounds climate crisis narratives and invites public engagement with AI-generated ecology.
Abstract
This paper discusses the artwork "Visions of Destruction", with a primary conceptual focus on the Anthropocene, which is communicated through audience interaction and generative AI as artistic research methods. Gaze-based interaction transitions the audience from mere observers to agents of landscape transformation, fostering a profound, on-the-edge engagement with pressing issues such as climate change and planetary destruction. The paper looks into early references of interactive art history that deploy eye-tracking as a method for audience interaction, and presents recent AI-aided artworks that demonstrate interactive latent space navigation.