Charles Voysey
Concept
I modelled a Charles Voysey building in Rhino, rendered multiple perspectives, from different angles, to get a complete overview of the buildings characteristics. A LoRA (Low-Rank Adaptation) was used to fine-tune a Stable Diffusion model on the rendered images. LoRA allows the creation of more personalized models without altering the entire model’s weights, making it efficient for transferring the specific style and form of the architectural features. Like this, the model was able to learn the architectural elements unique to Voysey’s design, such as the distinctive chimneys, steep gables, and cottage-like forms. This fine-tuned model was used to generate new architectural images, which maintain the core aesthetic principles of Voysey’s typology, while allowing the AI to generate variations or new interpretations of the form.
Training the model on Voysey’s architectural features, the AI learns what is essentially an associative form of aesthetic reasoning. The AI learns to associate specific forms with the principles of Voysey’s designs.
The model’s ability to synthesize new images based on learned aesthetic principles from the training data can be compared to cognitive processes, where individuals process and interpret architectural or artistic forms based on prior knowledge.
The experiment is a small contribution to the conversation on how modern tools (like Stable Diffusion) interact with historical and psychological frameworks of design. Thus, machine learning is used to explore and extend the design legacy of an Arts and Crafts, similar to the avant-garde exploration of aesthetics and form in the early 20th century. Voysey himself worked with associations between natural forms, vernacular building traditions, and a modern architectural language.
The experiment extends the associationist and experimental methodologies of the early 20th century avant-garde into a new computational and digital paradigm
Stable Diffusion „Cardboard“ Models
A
B
C
D
E
F
G
H
I
Re-associationg, pre-existing Forms
