Person:

Haridis, Alexandros

Recent advancements in multimodal Generative AI may democratize specialized architectural tasks like interpreting technical drawings and creating 3D CAD models which traditionally require expert knowledge. This paper presents a comparative evaluation study of two systems—GPT-4o and Claude 3.5—in the task of architectural 3D synthesis. It takes as a case study two buildings in Palladio’s Four Books of Architecture (1965): Villa Rotonda and Palazzo Porto. High-level architectural models and drawings of the buildings were prepared inspired by Palladio’s original text and drawing corpus. Through sequential text and image prompting, the study characterizes intrinsic abilities of the systems in (1) interpreting 2D/3D representations of buildings from drawings, (2) encoding the buildings into a CAD software script, and (3) self-improving based on outputs. While both systems successfully generate individual parts, they struggle to accurately assemble these parts into the desired spatial relationships, with Claude 3.5 showing overall better performance, especially in self-correcting its output. The study contributes to ongoing research on benchmarking the strengths and weaknesses of off-the-shelf AI systems in intelligent human tasks requiring discipline-specific knowledge. The results show the potential of language-enabled AI systems to act as collaborative technical assistants in the architectural design process.

This paper argues that changes or shifts in descriptions of architectural objects or elements have played a critical role in advancing architectural design across history. Description changes are possible when the elements are amenable to a multiplicity of interpretation, to shifts in their structural configuration, appearance, or stylistic meaning. An original compilation of examples is presented that illustrates such descriptive changes across history. These range from the formation of the Doric order to Romanesque, Gothic, Modern and Postmodern architecture. These descriptive changes are shown to have stimulated important generative processes, from forming the triglyph to crafting a Gothic pinnacle or spatially and structurally organizing Terragni's Danteum. The diversity of the examples demonstrates the relevance of descriptive change to all aspects of architecture: from purely structural to purely ornamental.

"Description changes" are represented as rule-based computational processes guided by identity rules, a class of shape rules developed within the design computing theory area of shape grammars. The compilation of the examples illustrates how visual calculating with identity rules can be applied to real architectural contexts. Due to the rising interest in data-driven machine learning, whether for classification or for generative modeling, current architectural design discourse often revolves around a need for creating digital datasets that statically represent a record of architectural knowledge (e.g., plans, three-dimensional geometries, building component libraries). By raising the importance of description changes that have characterized architectural innovation, this paper argues for adopting a new lens in the representation of architectural knowledge in data, namely, one that embraces the possibility and need for changes in descriptions of architectural elements and a multiplicity in their meanings.

We introduce JONES-19, a high-quality image dataset documenting 1,901 ornament designs belonging to nineteen human cultures. The images and their annotations are based on an open access archive of The Grammar of Ornament (London, 1856), by Owen Jones. The dataset poses numerous challenges as a benchmark for computer vision classification tasks and for research at the intersection of machine learning and art and design: a small sample size, image samples of human-designed artifacts rather than common objects in-context or natural scenes, imbalanced class distribution, and image distinctions based on fine details involving line patterns, reliefs, and colors. As a design-inspired dataset, JONES-19 can serve as a benchmark for various research fields, such as visual recognition, data-efficient learning, art-historical research, architectural style analysis, and cultural heritage. This paper describes the curation of the JONES-19 dataset and reports a baseline classification benchmark that evaluates the suitability of the dataset for training classifiers and exposes insights into inter-class relationships–particularly cultural similarities–by examining patterns in misclassification errors. The dataset and its accompanying documentation are available at: https://huggingface.co/datasets/harvardseas-cultural-ornaments/JONES-19.

Construction lines and registration marks in shape grammars ground the appearance of shapes to provide an algorithmic approach to visual calculating. Here, the two concepts are studied in unity as a new object called point-line arrangement. This chapter develops such topics as the comparison and classification of shapes by their arrangements, the characterization of the “geometries” that arrangements give rise to, the algebra of arrangements, and others. In this way, it provides a more holistic view of construction lines and registration marks, beyond the usual roles they receive in algorithmic implementations of shape grammars.

Following Kant’s view of drawing or shape as the “proper object” of aesthetic judgments in architecture, I present an interpretation of a central concept in his theory of aesthetic judgment, viz., the free play of imagination and understanding, as visual calculating in shape grammars. Calculating with identity rules formalizes Kant’s reflective judgments in free play, which he explains as imagination sustaining a “lively engagement” with form. This interpretation departs from determining judgments, which underlie twentieth-century mathematical and computational approaches to aesthetics. With this interpretation in place, I address a central issue concerning computation and aesthetic intelligence, engaging Kant’s concept of “adherent beauty”: How are we to employ computation as a practical method for value judgment while preserving free play’s reflective property that refreshes aesthetic experience, especially when creative work must meet defined functions and end-goals?

At the 1956 Dartmouth Summer Research Project, creation and evaluation processes were identified as one of seven key dimensions of human intelligence that future AI research should address. Nearly seventy years later, computational systems increasingly participate in processes of aesthetic judgment, generation, and transformation across architecture, art, and design. Beyond Data-Driven Aesthetics examines 20th- and early 21st-century computational aesthetic systems in architecture and the applied arts that formalize processes of creation and evaluation beyond purely data-driven approaches. Presented as a research exhibition at the Keller Gallery at the Massachusetts Institute of Technology, it brings together work from academic and industry contexts across the United States and Europe that treats computing as a medium for addressing foundational questions of aesthetics, including how design and art are judged, generated, and transformed. Drawing from contemporary design methods and technology studies on software reconstruction, physical making, and data visualization, the exhibition translates select aesthetic systems from archival sources and academic literature into tangible and experiential formats. In doing so, it extends traditional modes of academic scholarship and public communication into spatial, visual, and material form.

Design and architectural archives encode expert human knowledge in graphical formats, providing a critical testbed for design-inspired Machine Learning (ML) challenges absent with typical computer vision benchmarks. Building on JONES-19, a small-size image dataset based on The Grammar of Ornament (London, 1857), we evaluate the discriminative performance of Convolutional Neural Networks (CNNs) in two model training strategies: (a) ImageNet pretraining for domain-general “visual common sense,” and (b) learning from scratch on the design data in JONES-19. We find that while domain-general priors improve discriminative performance, learning from scratch augmented with repeated local sampling (multi-crop) effectively recovers these gains. For highly structured design data, local design-driven representations provide sufficient foundation for learning, challenging a reliance on massive general-purpose pretraining. These findings suggest that in specialized design domains, careful curation of smaller high-quality datasets that capture empirical and formal design principles may prove more effective and informative on the nature of a particular design domain than prioritizing large-scale data collection.