LUKE KRATSIOS

ARCHITECTURAL INTELLIGENCE:
HARNESSING Al FOR DESIGN INTENT

HOW CAN THE REAL-TIME MASKING AND LAYERING OF AI-GENERATED IMAGERY BE USED TO EFFECTIVELY REALIZE AND REFINE DESIGN INTENT?

LUKE KRATSIOS

This thesis investigates and develops state-of-the-art technologies in the field of generative content and examines how designers can leverage inherent training biases and input data to swiftly generate textures and imagery for use within the design process. 

Ultimately, the developed workflows aid in formulating a cohesive project vision from early stages. By combining proceduralization, simplified modeling techniques, and intelligent masking, the design process accelerates. This not only enables the creation of high-quality renders but more importantly, encourages a nuanced approach to design thinking. 

CLASS

CORNELL

B.ARCH 23

ADVISORS

CHRISTOPHER BATTAGLIA

PATRICK KASTNER

DONALD GREENBERG


TECHNICAL CONSULTANTS

JOHN WOLFORD

LEUL TESFAYE

RESEARCH HIGHLIGHTS

THE PROBLEM TEXTURE PRODUCTION MASKING AND LAYERING (NOT) THE BARCELONA PAVILION PARAMETRIC DESIGN DESIGN STUDY

According to industry standards in design representation, that number is around 5,000 dollars. Beyond the prohibitively high cost for most firms, this massive expense speaks volumes to how these renders are being utilized in our industry. These are not made for or by designers, they are purely marketing tools used to sell a product.

 

The quality provided by affordable and fast arch-vis alternatives falls short of yielding meaningful design feedback or freedom. They result in distracting and uninspiring depictions.

COMPLETE RESEARCH

GALLERY

COMPLETE RESEARCH

GALLERY

RESEARCH HIGHLIGHT

What if the designer of a project has no experience with complex 3D software and digital tools. Or they prefer to design with physical models in a tangible way. In any case, the goal is to model quickly and freely with less concern for craft and more of an emphasis on rapidly capturing design ideas in their head. How can the designer still make full use of this technology and material information as feedback towards design decisions?

SITE IMAGE

MODEL PHOTO

OVERLAY

GENERATED MAP

AI RENDER

By overlaying a photo captured on my phone of the physical model with the site image, then keying out the yellow background, the model can be placed in context. 

From there, a pair of images can be generated: a canny edge detection and an ai-generated depth map. This set of information, in tandem with the original color data, is all the information necessary to produce an final rendering.

GENERATED MAPS FROM MODEL PHOTO

EDGE DETECTION
DEPTH

MODEL PHOTO VS GENERATED RENDER

MODEL PHOTO
RENDER

DESIGN ITTERATIONS

MODEL ITTERATIONS

ACCOMPANYING RENDERS