Digital Media: Neural Bodies

This course explores generative artificial intelligence for volumetric, and especially architectural, modeling by considering the building as body from a computational and critical perspective. The friction between the building as a quasi-biological organism on the one hand and a precise geometric construct on the other has long been conceptually and formally productive. Moreover, the reciprocal interaction between global body forms and local patterns and features—such as animal patterns, clothing construction, or architectural discretizations—are at the core of parts-to-whole relationships that define morphology.  We will interrogate the reciprocity between architectural, biological, and mathematical form through the technical lens of datasets, AI, and geometry, as well as the thematic lens of fashion and garment construction. In particular, we will develop techniques that draw on hybrid architectural-biological datasets to generate speculative proto-architectures.

The course will look at both large architectural datasets—for example, beaux-arts plans—as well as large biological datasets—such as skeletal scans of comparative anatomy—and ask how the biological data might be understood architecturally and vice versa. Practices of anthropometry, zoometry, and surveying will provide one lens for such datasets. Beyond available image and 3d-scan datasets certain representation and imaging techniques, such as tomographic scanning, will be critical. Through AI techniques such as volumetric neural networks, students will explore bio-architectural generative spaces in a spatially sophisticated way.

The course will cover a combination of AI techniques and geometric methods. The first half of the class will introduce AI techniques of interest including image and latent space interpolation (Stylegan), text-to-image (DALL-E, Stable Diffusion) and text-to-model approaches (Dreamfusion, including 3d), and mesh stylization (CLIPMesh, Text2Mesh). In the second half, geometric methods based in Houdini will include surface discretization and processing, surface mapping, UV unrolling, and transformation, cloth folding and unfolding (e.g. Houdini clothing modeling), and kitbashing.

Students will develop three sequential exercises related to (1) pattern- and plan-making, (2) volumetric explorations of latent space and data-driven voxel forms, and (3) a quasi-architectural proposal that builds on the first two assignments and leverages the offered geometric techniques. Technically, students will also be introduced to many of the mathematical and computational concepts underlying neural networks, voxel-based modeling and image processing, and surface mapping. A key aspect of the class will be the development of productive workflows that leverage disparate tools for novel effect. Students will use AI tools like Google Colab, Tensorflow, Stylegan, Dreamfusion, and Shapenet as well as a range of surface modification, analysis, and discretization tools primarily in Houdini but supported by Meshmixer and Grasshopper. Students will also be introduced to certain procedural processes to discretize and resolve their proposals into printable or otherwise constructable forms.

The ambitions of the course extend beyond techniques of form making to critical perspectives on architecture as body and broader ideas of morphological visualization, analysis, and classification. The course thus engages ways of thinking about and measuring bodies and the surfaces that clothe them generally and considers these practices as constitutive of a rich mode of architectural production.