Harshdeep Arora

Advisor: Shawn Bailey

The image features three representations of a building design: a real-life photo with tinted windows and dark panels in a grassy setting, a digital rendering with simplified details in white and blue, and a physical scale model with corrugated vertical panels in off-white.
A multipanel layout showcasing architectural design through digital processes, models, and diffusion-based results.
Research Panel
A grid of eight 3D house models, varying in design, inspired by a black sketch.
Sketch Input to Design Variations through Latent Diffusion.
A node-based workflow in ComfyUI for Stable Diffusion, showing interconnected processes like image loading, upscaling, and saving. Nodes include "Load Image," "Apply ControlNet," and "Save Image," with preview images and colored data flow lines.
Stable Diffusion Workflow using ComfyUI node-based interface.
A comparison showing site surroundings for diffusion training (left) and potential diffusion results (right) with modern, geometric architectural designs.
Diffusion Model Training using Site Surroundings (left) showing potential results (right)
A tripartite image comparing a modern building's design: original (left), digital interpretation (center), and physical model (right).
Latent Diffusion on selected site(left) using trained LoRA, results interpreted into digital(middle) & physical(right) models.

Diffusing Design 

Developing Architecture Design through Latent Diffusion & Machine Learning 

Machine learning and diffusion have created a new path for creation. Achieving tasks through machine learnt systems, especially creative work for designers, opens new possibilities.  But the base data of diffusion trained on existing architecture & context questions the credibility & possibility of finding something New. The methodology to create through this process follows alternative unconventional trajectories. Also, developing architecture through diffusion has its challenges, to be identified, addressed & possibly solved.  

As diffusion process adds Communication involving language/text/speech & visual/image referencing, in addition to conventional methods of Click & Command, conveying and processing ideal results involves various steps. Just like verbal & visual conversation, communicating intent in diffusion is a two-way process requiring input/guidance, assessment, editing & re-communication.  

Hence, this research presents various processes/workflows in diffusion resembling/inspired by contemporary means of developing architecture, achieving a variety of designs. Exploring workflows in guiding diffusion, experimenting through Sketching, Referencing, Language, these workflows & experiments will make way towards enabling designers to guide, communicate & learn with creative machine learning, critiquing upon its use, limitations, and future possibilities in Architectural practice.   

But what can we achieve by using Diffusion? Experimenting methodologies inspired by conventional architectural means like sun/shadow study, massing, sketching, this research tries to iteratively design meaningful objects & spaces guided by constraints & parameters. Designing with Diffusion not only can create intent, but more, creating anomalies, artefacts inspirations unfamiliar to a Design to Draft means, creating opportunities to find and develop with the diffusion process. Combining these conventional techniques with diffusion, including adding site, context & culture opens unique opportunities for design, vision and sensitizing the diffusion process specific to its use case. Communicating and finding in this hybrid design process takes new paths employing convention & diffusion. Like a music conductor, my role as an architect in this study will be guiding these diffusions, adding & refining necessary conventions (programs, circulation, & others) creating wholesome sensitive architecture.