Augmented Intelligence for Hotel Design
By Lawrence Adams
Image generated from Midjourney for high-rise resort hotel.
I am old enough to remember when T-squares, triangles, and compasses were the primary tools architects used to design and document building design. In 1982, Autodesk released AutoCAD, which revolutionized the practice by taking architects off drafting boards and onto digital computer screens. Then, in 1987, Graphisoft released ArchiCAD, the first BIM (Building Information Modeling) platform, which took architects into the world of 3D digital modeling for building design and documentation.
In 2000, Revit replaced ArchiCAD in the US as the go-to tool that included advanced 3D modeling, parametric features, and building optimization algorithms. Parametric 3D modeling was significantly enhanced in 2008 with the addition of Grasshopper to the Rhino platform to advance modeling and visualization of freeform and complex geometries. Revit, in concert with a wide variety of Generative Design (GenD) platforms, continues to advance building optimization with computational design algorithms.
Until recently the focus of these tools has been on logical, computational and analytical processes that can be categorized as Generative Design supporting architects’ left-brain design capabilities for functional and efficient building design. The advancements in these platforms have freed up our workloads leaving more time for creative activities associated with artistic expression, innovation and inspirational work that we associate with right brain capabilities. While these GenD tools enhance our creative design methods and facilitate visioning of architectural design to fantastic levels, they have not, by themselves, produced actual creative design output. The emergence of Generative AI (GenAI) is offering an amazing leap forward as an effective design tool supporting architects’ production of creative and innovative design solutions.
Generative Artificial Intelligence (GenAI) is a subset of AI that refers to a class of algorithms and methodologies that use neural networks to learn patterns in data to create new and original content. Recent developments in GenAI have spawned new tools and methods for predictive algorithms that function much like human creative cognition. Their impact has been profound, enhancing many human artistic endeavors in music, writing, film, theater and art. Our interest here is in its broad impact on architectural design and through a narrower lens with hotel design. We will look here at emerging GenAI platforms that create novel architectural images, 3D models, floor plans and spatial composition.
2D Images
Three powerful GenAI tools used for creative image output were released almost simultaneously in 2021-22, namely Midjourney, DALL·E, and Stable Diffusion. Each platform’s format is unique but has similar algorithmic functionality that relies on carefully crafted text prompts directed at richly populated datasets to produce sensational creative output. GenAI image creation is revolutionizing architects’ capabilities for early-stage design ideation and visualization. As we gain more control over the process through prompt engineering, curated datasets, and reiterative workflow, we establish genuine authorship for style-based output that facilitates purposeful and innovative architectural design.
Two grand hotel lobby images generated with Midjourney based on similar text prompts but with distinctly different styles descriptions.
While novel image creation using text-to-image GenAI platforms has huge value as a tool informing our creative design process, it is limited to 2D imagery or snapshots taken from a single vantage point. The algorithm is fundamentally blind to the figurative dark side of the moon. Much work remains to take a pixel-based image and develop its form in 3D digital format that will lead to productive building design and documentation. GenAI algorithms that will produce full 3D digital models are evolving but have a long way to go before being useful as effective design tools.
3D Models
Initial attempts to resolve the “dark side” issue have been with multi-view generation, where predictive algorithms leverage multiple camera-poses to predict angular views to reveal the complete form of an object or building. MVDream and Zero123++ are two that are making advances in multi-view generation. Other platforms are gaining some success in producing full 3D digital models through generative diffusion algorithms. Lumen.ai, Trellis 3D, and Meshy AI are three GenAI platforms that have successfully generated novel 3D content, though their quality level remains sketchy and somewhat messy, and still leaves the architect with much work to do to bring it to a high level of formal accuracy and utility. The goal is to develop an algorithm that produces a comprehensive model with a substantial level of detail that, after iterative refinements, can be imported to a BIM platform and lead to fully formed architectural design and documentation.
3D mesh-based model created with Meshy AI. Large sprawling beachfront resort, 3-stories, Shingle Style, symmetrical layout, with large pool in front and canopy at the entrance, ornamental rooftop.
While there is optimism that platforms will soon be available as design tools facilitating the creation of unique and purposeful architectural form, when considering the complex creative process required in producing any form of architecture, we still need to push forward on two other essential and intrinsically connected components of the design process, i.e., imaginative floor planning and spatial composition.
If we set aside the more rational design processes that require left-brain problem solving and look at components that are fundamental to creative architectural design we find 3 basic legs to the stool: 1) Architectural Form, 2) Floor Plans and 3) Spatial Composition (Form, Plans and Space). Materials, finishes, lighting, color, ornamental details, furniture, and other visual features of architecture and interior design may be integral to the hotel guest’s experience and in many cases may even be primary over space and form, but we must always start with those basic 3 legs.
Floor Plans
As we architects begin to conceptualize design problems, after studying the program, we are likely to start by sketching floor plans. Solving complex floor plans is foundational in the composition of architectural design. There are numerous GenD platforms that facilitate logical functional floor planning. Examples include TestFit, LaiOut, and SPG (Space Plan Generator) which are math-based, formulaic programs that take functional relationships using matrices and bubble diagrams and produce optimized plan alternatives rated based on predetermined parameters and precepts. The (cookie-cutter) outcomes are generally practicable, albeit predictable and unimaginative.
Parafin is a computational model that produces efficient, logical hotel floor plans derived from well-defined brand standards and prototypes provided by major hotel chains. None of these computational platforms are equipped to infuse plans with distinctive style or to provoke imagination and delight. GenAI is only scratching the surface with platforms that produce original style-based plan solutions that delight us and inspire innovative architecture.
Finding meaningful patterns in floor plans for training neural networks requires comprehending the syntax of graphical representations of space. A floor plan may be read as a 2D graphical pattern, e.g., a Miesian plan has a minimalist non-axial orthogonal pattern, where an ornamental Baroque plan has curves, hierarchical geometry and formal arrangement of spaces in a cross-axial pattern. But beyond 2D graphical patterns, a floor plan in a dataset should have strong reference to the volumetric character it represents. Paring of floor plans with volumetric representation, in prompts and data, is important for a diffusion process toward viable novel floor plan creation.
Design style plays a pivotal role in floor plan synthesis with spatial composition. Stylistic representation of both components must be incorporated for any algorithm to produce valid outcomes. In his thesis, Stanislas Chaillou at Harvard Graduate School of Design writes, "Style carries a fundamental set of functional rules that defines a clear mechanic of space and controls the internal organization of the plan."
Chaillou’s Harvard research developed metrics of floor plan characteristics used as filters in the generation pipeline leading to novel plan solutions. It follows that semantic labels that define these qualities of floor plans may inform training mechanics for achieving creative floor plan content. Foster + Partners Architects with University of Leeds created a diffusion model that succeeded in unraveling the intricate nuances of design mutations within floor plans by understanding the dynamic interplay between the graphic depiction of design elements. Their experimental model generated entirely new configurations by learning from diverse building types and architectural design variations.
Spatial Composition
Compared with 3D object generation, 3D scene creation is more challenging due to the complexity and diversity of indoor scene geometry, which requires greater computational bandwidth. Recent advancements in video game design have had amazing results in 3D scene composition using programs like Unreal Engine, but much of that is hand constructed using 3D modeling algorithms like Grasshopper versus imaginative spatial compositions generated through diffusion modelling.
A number of new experimental GenAI platforms attempt to address the need for 3D scene generation to serve video gaming, filmmaking, VR/AR and Embodied AI (robotics, drones, autonomous vehicles). But most are limited to arranging 3D furniture models in preset room configurations and do not directly lead to creative spatial composition as imaginative 3D volumetric models. DiffInDScene and Text2Room are two platforms that have a broader application to creation of novel spatial composition compared with those limited to arranging objects within a predetermined spatial context.
Spatial Choreography
Asking GenAI to compose a sequence of spatial experiences is a tall order. While we are just scratching the surface of being able to generate 3D models of complex environments, stringing together a series of views that meaningfully deliver a focused spatial experience requires a higher level of neural spatial understanding. Such a composed spatial sequence might best be represented in some form of a video stream, though a procession of spaces could be graphically scored like a dance routine as demonstrated by Lawrence Halprin in his book, RSVP Cycles (1969).
Our experience of space is not static. It requires movement. We start with the proposition that the experience of moving through space can be shaped by design through spatial choreography. If we think of the initial hotel guest experience from arrival court, through the lobby, registration and circulation routes past the restaurant, bar, and up the elevators to the guestroom, we can imagine a composition of volumes along the path, each with its own characteristics (see First Impressions: Designing the Guest Arrival Experience , Hotel Business Review, 8.25.24, by Lawrence Adams). In John Portman’s design for the arrival sequence of the Atlanta Hyatt Hotel, the guest is spatially compressed and decompressed along the path from vehicle to guestroom, accentuating the dramatic spatial experience of that early atrium hotel.
The challenge is to link spatial compositions together in sequence to provide a full choreographed spatial experience. One promising tool for sequential spatial design is the Isovist Field, a graphic device that records metrics of spatial characteristics that can then be used through diffusion modelling to compose a novel sequence of spatial experiences. Scenescape by NVIDIA, is an experimental video platform that uses perpetual view generation of a spatial sequence by stringing together a series of 3D mesh-based models, so that each scene is prompted by the preceding scene.
Full detail 3D model of Chapelle du Crous interior from SketchFab. 3D models based on high-detail scans of interior environments provide a rich dataset in the pipeline for GenAI spatial composition.
Game Engines
Propelled by increasing popularity for video games, the game engine industry has greatly advanced 3D visualization in recent years and has achieved a level of sophistication rivaling many of the current architectural software and methodologies. Unreal Engine 5 (UE5), TwinMotion, and Houdini are platforms that excel in complex 3D scene creation. Unreal Engine is a powerful 3D computer graphics game engine developed by Epic Games as a comprehensive real-time 3D creation tool used primarily for video game development, but has also found applications in various industries such as film, virtual reality, and architecture. UE5 achieves brilliant cinematic quality level output using a virtualized micro-polygon system to generate highly detailed mesh geometry representing elaborate 3D scenes.
A game design workflow for scene creation with UE5 often includes the composition of pre-defined high-poly objects and assemblages from 3D libraries. These libraries provide a wide range of complex 3D models that are assembled into thematic scene elements providing a shortcut from time-consuming hand modeling of each element in a scene. Many of the assemblies and their parts are architectural in nature, containing buildings and building elements. These models could be utilized as a richly curated dataset for training diffusion models to create 3D objects, 3D spatial composition, and choreographed spatial orchestration.
Datasets
Data is a fundamental component of any AI platform and is the foundation upon which Generative AI functions. The quality, diversity, and volume of data that a GenAI model is trained on is a primary determinate of the quality level of its creative output.
A GenAI process starts with a verbal or visual prompt that is entered into the algorithm to guide the output. But a prompt will not be effective unless the dataset that trains the diffusion model is aligned with the prompt’s intentions. Prompt engineering is a term that refers to skillfully composed text descriptions to guide the model towards an intended outcome. Equally important is data engineering which refers to the development of robust, high-quality, consistent datasets for training the generative model. The alignment of prompt to data depends heavily on the syntax common to both. To facilitate alignment, data samples need to be skillfully annotated. GenAI output relies on well-crafted, detailed and accurate annotations of image or three-dimensional data. Paring images with text labels, captions and annotations is important to correctly instill the model with complex relationships between visual elements and syntactical descriptions.
GenAI platforms that generate novel 2D imagery, access vast quantities of imagery on the internet. LAION-5B is the largest freely available 2D dataset of its kind, with over 5.8 billion image-text pairs. Objaverse-XL, by Paul Allen’s Allen Institute for AI, has the most comprehensive digital 3D collection to date, with over 10 million unique 3D objects. A wide variety of floor plan images exist throughout the internet and are included with countless 2D image datasets. However, nearly all the publicly accessible floor plan datasets are residential without representation of other building types.
RPLAN is an example of a large dataset of residential building floor plans containing 80,000 annotated floor plan images but is limited to the genre of modern urban housing layouts. Datasets populated with thousands of floor plans typologically catalogued would allow the diffusion process to identify patterns and relationships and create new plans predictively. ScanNet and HM3dSem are two richly annotated datasets of high-res volumetric scans that provide a wide range of commercial and civic spaces supporting 3D spatial diffusion models.
Game Design Libraries
Game design workflow for scene creation with platforms like UE5 usually includes the composition of pre-defined high-poly objects and assemblages from 3D libraries such as Turbosquid, Sketchfab, and Kitbash3D. The Kitbash3D library provides a wide range of complex 3D models that are assembled into thematic scene constructions, providing a shortcut from time-consuming hand modeling of each element in a scene. Kitbash offers highly detailed preassembled thematic kits created by artists and architects as stylistic components for depicting cinematic scenes. Thematic collections from ancient temples to futuristic cities include titles such as Diesel Punk, Lunar Base, Age of Egypt, and Atlantis.
Kitbash libraries currently have over 1,000 individual 3D building models in a wide variety of highly imaginative styles. These collections could be utilized as richly curated datasets for training diffusion models to create 3D architecture, spatial composition, and spatial orchestration. Pairing these 3D assets with stylistic annotations in a curated dataset would offer significant potential for diffusion model generation of 3D architecture and scene composition.
The Kitbash3D BRUTALIST DYSTOPIA kit features a collection of cyberpunk-style 3D architectural assets that could populate a dataset for 3D building design generation.
Curation of data is itself an artform. Selecting and annotating data for composing a curated dataset provides the boundaries necessary for effective and purposeful creative output. Curating data is the art of setting boundaries that encourage imaginative output with GenAI. Many users of AI create proprietary datasets that promote a bias toward a particular or personal style of architecture, that avoids plagiarism, and result in genuine authorship of the output.
Synchronized Design
In architectural design we rely on our left-brain intelligence by using logic, math, science, rational thought, and systematic analysis to solve design problems of building technology and functional planning. But to deliver beauty and experiential delight we must employ our right-brain intelligence for imagination, artistic expression, sensuality, and a sense of adventure. True design innovation requires a balance of the two to create beautiful yet functional buildings. Synchronizing left brain capabilities of GenD with right brain capabilities of GenAI is the next big hurdle in the profession’s embrace of AI.
Concluding Thoughts
Where do we want to be in 10 years? Will GenAI platforms for 3D models, floor plans and spatial composition be part of our everyday arsenal as essential tools for designing hotels and other buildings? Will GenD and GenAI be synchronized, so we arrive at a highly functional, optimally efficient, and sustainable building that is beautiful, inspiring, and innovative? In 10 years from now, will it be possible to have AI produce a fully formed, accurately detailed, code-compliant, set of design and construction documents automatically with minimal guidance from the architect?
While we may find these scenarios both exciting and frightening, we need to understand the possibilities and never lose sight that AI is a tool, another arrow in our quiver, that requires guardrails and controls to maximize its effectiveness in advancing the profession.
"You are not going to lose your job to an AI, but you are going to lose your job to someone who uses AI."
- Jesen Huang, CEO Nvidia
Mr. Adams
HotelExecutive retains the copyright to all articles published on HotelExecutive.com. Articles cannot be republished without prior written consent by HotelExecutive.
