<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Yuhan Zhang on Medium]]></title>
        <description><![CDATA[Stories by Yuhan Zhang on Medium]]></description>
        <link>https://medium.com/@zhangyuhan122333?source=rss-c73dd5f7d370------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*qp9ttKjNi01PWxl4.</url>
            <title>Stories by Yuhan Zhang on Medium</title>
            <link>https://medium.com/@zhangyuhan122333?source=rss-c73dd5f7d370------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 17:50:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zhangyuhan122333/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Architectural Facade in the Eyes of AI]]></title>
            <link>https://medium.com/@zhangyuhan122333/architectural-facade-in-the-eyes-of-ai-24dcee7575d5?source=rss-c73dd5f7d370------2</link>
            <guid isPermaLink="false">https://medium.com/p/24dcee7575d5</guid>
            <dc:creator><![CDATA[Yuhan Zhang]]></dc:creator>
            <pubDate>Mon, 10 Jan 2022 20:16:53 GMT</pubDate>
            <atom:updated>2022-01-10T20:16:53.253Z</atom:updated>
            <content:encoded><![CDATA[<h4>Exploring Facade as an Artifact Processed and Generated by Computer Vision</h4><p><em>Yuhan Zhang, Yale School of Architecture | Spring 2020</em></p><p><strong>Abstract</strong></p><p>The architectural façade is one of the most expressive elements of architecture. It constantly evolves with technologies and architectural theories. In this article, I explored recent machine learning algorithms relating to computer vision and image processing, and investigated how architectural façade responds to these technologies. By feeding a previously trained 2D façade model with manipulated 3D imageries, I blurred the vision of AI in terms of reading the façade input and received different outcomes comparing with the original facades. The article also discusses the future possibility for the façade to evolve with artificial intelligence and generative design.</p><p><strong>Introduction</strong></p><p>The architectural facade has been one of the most important elements of architecture since the establishment of architecture as a subject and profession. It is the first thing that people see when they approach an architecture, and where the experience of architecture starts. In the dictionary, its first definition is “the front of a building”, while it also has a second definition meaning “a false, superficial, or artificial appearance or effect”. The façade of architecture can reveal many things of the architectural space behind: its style, its function, its composition, its spatial quality and order, although it is not always an honest presentation. Architectural façade has always been evolving with architecture, and has been established with important theories that categorize and explains the architecture behind. The change of the façade system is closely related to the revolution of technology and philosophy development of human civilization. As we are at the turning point for artificial intelligence as the new era of human technology, how would architecture evolve with AI and how would architectural façade embody that evolution?</p><p>To begin with a historical background on architectural façade, I will discuss them in the age of high renaissance, modernism and post-modernism. High renaissance architecture is characterized by symmetry and proportion, which is illustrated both in its plan, space and facade. Renaissance façades are symmetrical around their vertical axis. For instance, church façades of this period are generally surmounted by a pediment and organized by a system of pilasters, arches, and entablatures. The columns and windows are organized with a progression towards the center. As they are mainly religious buildings, the order on the façade are intended to show the order of space inside, and in turn the hierarchy of religion it’s representing. The ornament functions not only in form-making, creating depth and foreground, but also as an embodiment catholic storytelling and ethics. As these architectures mainly to be looked at from one direction, façade remains one of the most important component and serves an indispensable role for what the architecture wants to achieve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/462/1*LbfdP3jqPmBruGJiv0dn4A.png" /><figcaption>San Giorgio Maggiore, Venice</figcaption></figure><p>In the age of modernism, architecture began to abandon ornament and embrace an honesty with structure and space. Le Corbusier made the five points of architecture in his manifesto towards the modern age, and he especially made a point on the free design of the façade — separating the exterior wall of the building from its structural function, which is supported by the steel/ concrete technology emerged at that time. The free façade rule allows him to establish the continuous horizontal window shown in Villa Savoye, which cuts the façade along its entire length and bring skylights into each room equally. From here, architectural façade began to show independence from its internal space and is less bounded by structural constraints. Despite this freedom, most modern architects still pursued a unification between the outer façade and inner space, to achieve the aesthetic of modern architecture as a “Gesamtkunstwerk”(total work of art).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/546/1*gK4Bfelq1hZKVcke-Qj9Pw.png" /><figcaption>Villa Savoye, Poissy</figcaption></figure><p>During the post-modernism movement, the pursuit of total work of art has been challenged and deconstructed. Architects began to examine the pure existence of each architectural element, and fundamentally question the idea of style and appropriateness. It is difficult to summarize postmodernism with one specific architect when the theories began to diverge into different streams. I will mainly discuss the beginning of postmodernism from Robert Venturi’s theory and Mark Foster Gage’s work as of postmodernism in current time. Mark Foster Gage’s work focused on Object-Oriented Ontology, which reaffirms the importance of architecture for its own sake and rich in complexity, rather than as a reductive mechanism for solving functional problems. (Gage 2015, 95) By excavating the meaning behind the very existence of everyday objects, it questions if architecture is also just such an object with its own meaning behind its form and physical existence. The façades of Gages’ design often tend to be a highly complex compilation with ornamental, non-abstract patterns and objects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/378/1*K2SHeD81fOWoOP-nx3ID8A.jpeg" /><figcaption>Guggenheim Museum Proposal, Helsinki (source:mfga.com)</figcaption></figure><p>In Robert Venturi’s view, what has changed for architects after modernism is the medium, and the façade of architecture is an excellent embodiment for it. The field of architecture has been heavily dependent on media, as architects produce drawings on paper to clarify and sell their idea. Even architecture itself is a media that architects works on, rendering a specific sense and thought of space and ambient designed by the architect. However, when the medium began to transition from three-dimensional space to two-dimensional displaying of information, the architectural facade is now capable of displaying information that’s beyond the architectural space behind it. It has become a big screen of display, which has accelerated the isolation between facade and architecture, leaning more towards the idea of “decorated shed” in Venturi’s definition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*BMfFqJQEi2Nx1L15mEIgOA.jpeg" /><figcaption>Robert Venturi — Duck vs. Decorated Shed</figcaption></figure><p>In his own work, Venturi put a “fake” antenna on the top of Guild House facade as a pure decoration, as he contends this unconnected antenna “would do the real work of attracting tenants by providing an image of a connected building”. While Guild House uses the antenna as an artifact on the façade to convey the architect’s message, the unbuilt work of Venturi — National Collegiate Football Hall of Fame — set precedent for treating the whole façade of building as an artifact of displaying information, foreseeing the flourish of LED screen commercial facade in our age.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/356/1*W5gGOVPuKP_fxK2CVyJEwQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*ujjdK_LuUqXiOizH5ziJHA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*gAl6MtvQsOJXmMZk_OuzPA.jpeg" /><figcaption>Postmodernism facade (Guild House, National Collegiate Football Hall of Fame, Times Square)</figcaption></figure><p>In our generation of architects, the evolution of the façade system is closely related to parametric design, a process that is based on algorithmic thinking which enables the expression of parameters in computer to the physical form of design. It allows for more complex geometries, carefully controlled change of details, and a new round of exploration for façade materials and representation. However, till today parametric design still exists in certain large-scale projects and has not penetrated through the whole industry, and we are already at another point of technology revolution of artificial intelligence. A new chapter of computer-aided design has began and still awaits exploration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*9RpXiij_IzUFbG7xivx3nA.jpeg" /><figcaption>Morpheus hotel in Macau designed by Zaha Hadid Architects (source: dezeen.com)</figcaption></figure><p><strong>Methods</strong></p><p>The method section includes a series of operations I did with machine-learning tools to manipulate architectural façade, trying to decipher how AI reads the information on a processed façade image and what they could generate out of the façade input. I started by exploring the possibility of transforming a 2D façade image into a 3D architectural space through computer vision, but later gradually transitioned to tempering AI vision with human-manipulated image, blurring the boundary of 2D and 3D in the eyes of AI and exploring the implied possibilities of architectural façade in the future.</p><ul><li>Reimaging façade of different periods in Google DeepDream</li></ul><p>DeepDream is a computer vision program created by Google engineer Alexander Mordvintsev which uses a convolutional neural network to find and enhance patterns in images via algorithmic pareidolia, thus creating a dream-like hallucinogenic appearance in the deliberately over-processed images.</p><p>I selected three façade images representing the three different time periods, movements or styles of architecture I discussed above as the input image for a model that is pre-trained with hundreds of dog images online.</p><p>The first image is the façade of San Giorgio Maggiore church in Venice, Italy. The neural network twisted the horizontal elements on the façade and thus creating arches. Originally flat marbles are contextualized with hologram colors, and when zoom in, high-level detailed ornaments begin to appear. The second image is the most representational façade of modern architecture — Villa Savoye. After DeepDream processing, the villa is full of arches and dogs. The pure white concrete façade pursued by modernism ideal now becomes ornamental marbles with excess of details, and a world of illusion and fantasy. It is not modern anymore. When I put the elevation image of Mark Gage’s Guggenheim Museum Proposal into the algorithm, the computer began to read all the 3D objects alchemized on the façade, and the stacks of daily random objects are unified with dog elements. Coloring of the marble made the already fleshy texture even more fleshy. A dystopian monster/ slaughterhouse has appeared in the eyes of artificial intelligence. By far postmodern is still postmodern, and this is the façade that has been read most successfully by the AI vision in the test.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/444/1*yQhgavgSpizYsEf1PQZb9w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/1*Te2SLs51ShFWcMBq6qBPAw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*KlKDuRBjglgWKQj0iJkC7A.png" /><figcaption>Google DeepDream result (source: author)</figcaption></figure><ul><li>Super flat art blended with three-dimensional architecture through Neural style transfer</li></ul><p>Then I tried the Neural style transfer tool in Google Colab, which is an optimization technique that uses deep learning to compose one image in the style of another image. I used the three façade images as the content images and three different artworks as the style images, and the result is blending them together though a trained model of online artwork images. However, the result of the neural style transfer is not as successful as my experiment in DeepDream. The blending of style mainly happened in colors as the input style images are super flat modern art of color blocks, the AI fails to read it as anything transferrable to the much more contextual façade images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZIhZpB7T1mrtmLLZTNrfIg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-_ulbUL1OsO5ul6xhIIqOQ.png" /><figcaption>Neural Style Transfer result (source: author)</figcaption></figure><ul><li>Style GAN — RunwayML test</li></ul><p>After the two tests focused on representational images of different architectural style. I began to use the CMP Facade Database, which is created by the Center for Machine Perception at the Czech Technical University in Prague. The dataset includes 606 rectified images of facades from various sources, which are from different cities around the world with diverse architectural styles. The dataset comes with two different groups of images: the first group is the actual photographs of the façades, the second group is processed color blocks representing different elements on the façade. Interested in the simplified and abstracted form of the facades, I started a test in RunwayML (a software platform to train machine learning models) to explore the representation of these color blocks using a Style GAN model, which is developed by Karras, Laine and Aila in 2019. It combines the GAN framework with style transfer, and is originally used to generate human facial images. RunwayML produced a latent space walking video of the façade dataset showing how they could transform from one to another and without any static form.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/256/1*vOyS5YmavrjMsX-fcw2rKg.gif" /><figcaption>Training Image Experiment — Latent Space Walk (RunwayML) (source: author)</figcaption></figure><ul><li>Transforming 2D façade to 3D objects</li></ul><p>After exploring how the computer vision is reading the façade data, I began to look into different methods that could potentially turn the 2D flat image into a 3D architectural space. I first started a test in Rhinoceros, a 3D modeling program commonly used by architects and designers. I created a 10 by 10 grid cube with pixelated 3D space, and then applied four selected façade images from the CMP façade data to each side of the cube. Then I began a series of extrusion and subtraction move by interpreting the colors into different parameters, resulting a Boolean calculated cube with a space constrained by four different facades. Then I also used the command Heightfield in Rhino, which created NURBS surfaces based on grayscale values of the colors in the image files given.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NrKmtfTPWkW_OnkgYo1NkQ.png" /><figcaption>source (CMP database)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*Kh2PUSFpPLxsj7cLBcqViw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nQVGPvFgWABWvOysHGVf5g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*dfouJN8wItkVZKJyPjCs9g.png" /><figcaption>Different operations in Rhinoceros in an attempt to transform 2D to 3D (source: author)</figcaption></figure><p>Some other 2D to 3D explorations I found outside the field of architecture are also very inspiring. In Learning Shape Abstractions, Tulsiani et al. presented a learning framework for abstracting complex shapes by learning to assemble objects using 3D volumetric primitives. It generates simple and geometrically interpretable explanations of more complex 3D objects. Gadelha at al. transformed 2D projections into pixelated 3D objects through “projective generative adversarial networks” (PrGANs) which trains a deep generative model of 3D shapes whose projections match the distributions of the input 2D views.</p><ul><li>Tempering AI vision of manipulated façade images — pix2pix</li></ul><p>Another example for decoding images and substantiate it with new information is an algorithm called Pix2pix, developed by Isola et al. in 2017 that investigated conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. Below are some examples for how they used it for different datasets and the outcome.</p><p>Inspired by their aerial-to-map and labels-to-façade generation, I saw another opportunity of giving depth to the flat 2D images by generating textures and shadows. When computer began to generate fake 3D projections out of simple labels in 2D, the space behind it is open for imagination and interpretation. In the last experiment, I used an interdisciplinary process combing the Height field operation in Rhinoceros and pix2pix on Google Colab to generate AI interpreted facade image with a manipulated facade database. The new input for the database is a plan-view with shadows from the Height-field surface I generated from CMP façade color blocks, which is already 3D object mimicking a 2D label. And then I used the model trained on the original CMP façade database to process the Height-field images, and compared the result with the original outcomes.</p><p><strong>Result</strong></p><p>The resulting surface object from Rhinoceros is drastically changed from the original façade color labels by letting the computer read the greyscale of each pixel as a parameter for extrusion. By combining these surfaces to the 10 by 10 cube, a complex form of space with vibrant colors appeared as a half manually transformed 3D object. Taking these non-flat surfaces with shadows as input images into pix2pix, I got different façade outcome than the original ones. Comparing my outcome with the original outcom es from pix2pix, I observed the algorithm still recognized the main elements from the color blocks and gave its own translation of it into actual façade components. However, details began to be blurred and more shadows are created. For instance, in the first group of comparison, it seems to emerge four more windows in the middle of facade than the original one. The orthogonal elements, both vertical and horizontal, are weakened to become a more fluid surface. In the second comparison, the images of my outcome seem more popping out in three-dimension as it has a more gradual change in the shadow profile, and each window has a slightly distorted eave. In a way, the programmed computer vision to read the flat façade surface is “messed” with human interventions of a pre-treated 3D image, creating a second message of faking in the final imagery.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3kKqnDJJZPxlVGo3KWc77w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*a2MKnvKydBpbuFlVGzKvgw.png" /><figcaption>Height-field Facade Transformation to 3D in Rhino (source: author)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_6uLRgtgXNDnj-FN6Yqgsw.jpeg" /></figure><p><strong>Discussion</strong></p><ul><li>Deep Learning of Deep Faking?</li></ul><p>With the current technology of machine learning to deal with imageries, computer could generate a complete set of speeches with fake facial movement and voices of a celebrity that people could hardly tell if they are true or not. Similar with the image processing I have experimented in this article, instead of generating meaningful content of the traditional architectural thinking, the results tend to superficial without a clear conscious or goal. The algorithms are good at “faking” contents by learning from the images in the trained model. However, the “faking” quality of the process might not be a completely negative thing to façade generation, as the secondary meaning of façade also indicates a “false, superficial, or artificial appearance or effect”. The façade generated by AI could become its own artifact rather than a truthful representation of the architecture.</p><ul><li>What is really a computer-designed façade?</li></ul><p>As we discussed at the end of the introduction, currently parametric design is the main approach for computer-aided design, but it is not a machine-generative process; architects behind it still takes full credits for the design. Artificial Intelligence opens for more possibilities in generative design and it could actually generate a product without human agency. However, the model is still trained by people, by selecting the algorithm and dataset, it is impossible to eliminate human bias. Moreover, creativity is the core of design, and the nature of machine learning is almost a counterargument for it as it does not really create anything “new”. So in the end, what would a computer-designed façade really be like? And how important would architects be at that point?</p><ul><li>What tools might appear for architects?</li></ul><p>The 2D to 3D transformation is an integral part of architects’ workflow: architects constantly switch between 2D and 3D in order to better communicate with clients and contractors. If AI could read the process and generate automatic transformation between the dimensions, architects’ workflow would become much more efficient. By bringing AI into the transformation and design process, there is also a chance for a combination of human vision and computer vision, which could fundamentally change architects’ perception of space, aesthetics, and the meaning of architecture. So, what’s the next evolution for architectural façade in the age of artificial intelligence?</p><p><strong>Bibliography</strong></p><p>Gadelha, Matheus, Subhransu Maji, and Rui Wang. 2016. “3D Shape Induction from 2D Views of Multiple Objects.” <em>ArXiv:1612.05872 [Cs]</em>, December. <a href="http://arxiv.org/abs/1612.05872">http://arxiv.org/abs/1612.05872</a>.</p><p>Gage, Mark Foster. 2015. “Killing Simplicity: Object-Oriented Philosophy In Architecture.” <em>Log</em>, no. 33: 95–106.</p><p>Isola, Phillip, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. 2018. “Image-to-Image Translation with Conditional Adversarial Networks.” <em>ArXiv:1611.07004 [Cs]</em>, November. <a href="http://arxiv.org/abs/1611.07004">http://arxiv.org/abs/1611.07004</a>.</p><p>Karras, Tero, Samuli Laine, and Timo Aila. 2019. “A Style-Based Generator Architecture for Generative Adversarial Networks.” <em>ArXiv:1812.04948 [Cs, Stat]</em>, March. <a href="http://arxiv.org/abs/1812.04948">http://arxiv.org/abs/1812.04948</a>.</p><p>Mordvintsev, Alexander, Christopher Olah, and Mike Tyka. “Deepdream-a code example for visualizing neural networks.” <em>Google Research</em> 2, no. 5 (2015).</p><p>Shao, Grace. 2019. “What ‘deepfakes’ Are and How They May Be Dangerous.” CNBC. October 14, 2019. <a href="https://www.cnbc.com/2019/10/14/what-is-deepfake-and-how-it-might-be-dangerous.html">https://www.cnbc.com/2019/10/14/what-is-deepfake-and-how-it-might-be-dangerous.html</a>.</p><p>Tulsiani, Shubham, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. 2018. “Learning Shape Abstractions by Assembling Volumetric Primitives.” <em>ArXiv:1612.00404 [Cs]</em>, August. <a href="http://arxiv.org/abs/1612.00404">http://arxiv.org/abs/1612.00404</a>.</p><p>Venturi, Robert, Denise Scott Brown, and Steven Izenour. 2017. <em>Learning from Las Vegas</em>. MIT Press.</p><p>Venturi, Robert, Martino Stierli, Museum of Modern Art (New York N.Y.), and David Bruce Brownlee. 1977. <em>Complexity and Contradiction in Architecture</em>. The Museum of Modern Art.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=24dcee7575d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>