Digital Tools, Learning and AI

Lighting design has historically been treated as a technical discipline focused on visual performance and energy efficiency. Yet contemporary buildings are increasingly populated by sensing technologies, connected infrastructure, and intelligent control systems that fundamentally alter how space can interact with its occupants. This paper argues that lighting can no longer be understood as an isolated design layer, but must instead be considered part of a broader sensory architecture that perceives, interprets, and responds to human presence.

The study introduces the concept of Occupant Importance, a design philosophy that positions the occupant—not the building system—as the central organizing principle of spatial environments. Within this framework, lighting, automation, security, networking, and audio-visual systems operate as an integrated sensory infrastructure that continuously senses environmental and behavioural cues to enhance comfort, safety, wellbeing, and energy performance.

The paper further explores how emerging agent-based artificial intelligence can transform buildings from passive infrastructures into adaptive environments capable of learning occupant patterns and autonomously optimizing spatial conditions. In such environments, lighting functions simultaneously as illumination, sensing interface, and communication medium between humans and intelligent building systems.

The research proposes that the future of architecture will move toward autonomous spatial ecosystems, where integrated sensory technologies enable buildings to dynamically adapt to human behaviour while advancing long-term sustainability and lifecycle performance. This shift reframes lighting design from the provision of light to the creation of responsive environments centred on human experience.

Introduction

Lighting plays a critical role in architecture, industrial design, urban planning, and lighting design, influencing how environments and objects are perceived. Visualization methods vary widely. From laboratory reconstructions of lighting setups and building interiors to software-based simulations (e.g., Revit, Rhino), which can be costly, time-consuming, and require a lot of data, to approximations in modeling tools such as Blender, Keyshot, or Maya, which provide simpler visuals with reduced accuracy.

Recent advances in generative deep learning have expanded the possibilities for realistic neural rendering, particularly in relighting images of objects or full environments. Research also explores relighting 3D scenes and meshes using diffusion models, further bridging the gap between simulation and real-world lighting effects. Visuals and Examples: https://anonymous.4open.science/r/Neural-Relighting/README.md

Methods

We propose a similar approach to the state-of-the-art by utilizing a deep learning model for 2D image relighting as part of a 3d rendering visualization, where the calculated screen space depth and normal views are used as input to the model, together with a camera rendering of the object, the lighting direction, and a prompt explaining the lighting setup. This way, designers can easily visualize and change the lighting conditions for 3D models.

As a basis for the prototype, we use the IC-Light model, which is normally used for 2D image relighting. To fit the model as part of a rendering pipeline, we built a 3D visualization application using PyVista. It provides basic interaction controls to the user – rotating the camera, zooming, and panning.

Our proposed pipeline takes a 3D model and extracts three rendering views from the point of view of the camera. First, it renders an RGB rasterized image of the model with flat lighting to minimize the influence of any shadows on the generative model. Then it renders a normalized depth view, representing the distances of each pixel to the camera view. Finally, it renders a screen-space normal view, based on the calculated and smoothed normals of the 3D model, that provides an overview of the smaller surface details of the model. We then generate a gradient grayscale image representing the light direction. We combine the depth and normal views, together with the gradient, to produce a geometry-conditioned light gradient that respects both the 3D shape and the finer details of the model. We feed the RGB rasterized view and the geometry-conditioned light gradient to the IC-Light model. Finally, based on the text prompt given to IC-Light, a different relit view is generated. To provide a more interactive experience, we do the relighting part of the pipeline every time the user stops interacting with the viewer. The rendering and relighting take between 2 and 3 seconds, making it not real-time, but quick enough that multiple views can be rendered and explored in a short amount of time.

We test the application with 3D objects of different quality, complexity, and number of vertices. We test on 10 objects, like buildings, vehicle models, interiors, and single objects. We show visually the results, and the average time-to-relight from 10 different camera positions is also calculated.

Conclusion and Future Work

The current prototype generated a new relit view in around 2.72 seconds, making it not real-time, but very useful for prototyping and quick visualizations. The next step in our research will be user-testing the prototype with designers and architects, to better understand how this can be made into a useful part of their work pipelines. In addition, we combine the current implementation with pre-made HDR light maps for better adherence to specific lighting conditions, which is currently not possible with only gradients and textual prompts.

Generative AI tools are rapidly entering architectural lighting design workflows, promising efficiency gains and expanded creative possibilities. This presentation pivots from the hype about AI applications and focuses on the broader implications for lighting design as a profession. Rather than demonstrating the potential of these tools, participants will be introduced to a set of questions that address ethical implications of AI integration into lighting design practice that should be considered before adoption becomes widespread.

Several overlapping and varied ethical considerations will be shared:

Content Accuracy and Competence. Generative AI systems produce plausible but frequently inaccurate response including hallucinations and sycophantic pandering. How will our community maintain competence and professional responsibility when delegating tasks to systems prone to errors?

Environmental Impact. The computational and infrastructural costs of generative AI are unprecedented with expanded e-waste, energy consumption and water usage for data center cooling carrying substantial environmental consequences. This tension is particularly acute for the lighting design profession, which has prioritized environmental sustainability concerns. Adopting tools with significant negative environmental footprints presents a fundamental contradiction: How will we balance Generative AI use with the resulting environmental cost?

Data Privacy and Confidentiality. Free tools offer minimal data protection and may infer proprietary understanding from project information. How can the lighting design community contribute to data models in a more intentional and fair manner?

Intellectual Property and Authorship. Who owns AI-generated design outputs? What disclosure obligations exist? How can we correct the pattern of Generative AI systems being trained on copyrighted design work without creator consent or compensation in the future?

Inequity and Access Disparities. Automation will displace entry-level positions while concentrating benefits among organizations with resources to effectively implement these tools. What are the profession’s responsibilities to emerging practitioners and historically marginalized communities?

Workforce Development and Learning Loss. Cognitive offloading through AI delegation may undermine the deep learning and critical thinking that develops experienced designers. How do we preserve the learning pathways and address the structural changes required to build professional expertise?

Consolidated Wealth and Power. Reliance on proprietary AI infrastructure creates dependencies on a small number of technology companies and raises questions about professional autonomy.

Professional Conduct Standards. Existing fields like law and engineering have begun implementing professional ethics frameworks that address competence, communication, confidentiality, disclosure and supervision. What are next steps for the architectural lighting design community to establish similar forms of ethical guidance?

These ethics dimensions do not exist in isolation. This presentation will demonstrate how these ethical considerations overlap and reinforce one another, creating a complex landscape where decisions made in one area have cascading implications across the profession.

Stakeholders throughout the lighting design ecosystem will be impacted by Generative AI use in varying ways depending on their roles. This presentation include summary of findings from previous AI Ethics interactive workshops that illuminate where consensus exists and where disagreement persists.

Learning Outcomes

Identify key ethical considerations specific to generative AI in lighting design practice

Recognize gaps in existing professional ethics frameworks as applied to AI tools

Understand how ethical implications overlap and create cascading consequences across the profession

Consider how ongoing adoption of Generative AI tools at a personal and systematic might integrate ethical considerations

Learning and working environments are evolving increasingly from purely physical ‘real world’ environments to a wide variety and combination of digital, virtual and hybrid configurations, including ‘mixed-reality’ technologies and setups. This achieves differing degrees of immersive learning experiences. While students, teachers, and researchers work across these varying environments, the emotional, cognitive and perceptual differences exhibited by students working in these environments have not been systematically studied. Related research is lagging behind; emerging studies typically compare learning performance between physical-spaces and virtual reality environments using headsets, placing the individual into isolated learning environments. This leaves behind a significant portion of digital-physical, room-oriented educational environments. This project brings together expertise from three different disciplines – cognitive science, lighting design and architecture – to study the characteristics and impacts of room-oriented immersive digital learning environments on human stress-related responses and cognition.

A pilot experiment was conducted with participants to compare differences in their responses as they perform tasks in a physical, daylit educational space versus its room-oriented virtual simulation that used large-scale projections simulating the aforementioned physical space at 1-1 scale. The ‘real world’ physical learning environment in this experiment was a classroom at an academic institution on a 4th Floor South-East exposure allowing in ample natural light. Privacy screens were installed on the windows to allow daylight without drawing attention by possible changes outside the windows. The same space was modelled and simulated to scale in an immersive 12 x 10 meters physical-digital (‘room-oriented’) virtual environment laboratory with a 360 degree projection screen. Both spaces were set up with the same number and spacing of physical tables and chairs.

During the study, in both environments, up to eight participants at a time first habituate to the conditions in an adjacent space, are provided with a short introduction and provide informed consent. The volunteers are participating in small groups to mimic a small seminar type-class. Then, they enter the classroom and each sit at their own table to complete several attention and creativity tasks as well as a small survey. Light loggers (illuminance and spectrum) are worn around the neck as a pendant and are also placed on tables. In addition, wrist watches are logging biometric information, and saliva is collected at the beginning and the end of the study..

The aim of the study is to collect and compare participant’s stress responses (saliva cortisol levels, heart rate variability, and self-reported stress) and cognitive performance (attention and creativity tests) for both the real-life classroom setting and the simulated environment. The study is driven by the question of whether students will have similar stress responses and cognitive performance in the ‘real world’ classroom and the physical-digital simulation.

While the environments were matched in size, and adjustments were made in both environments to match, as much as possible, the visual and lighting qualities in both spaces, exact mimicking is not possible in the virtual simulation even with high-end projector and spectrally-tuning lighting equipment.

The study is performed in the afternoon between 14:00 – 16:00 with no electric light added in the actual classroom. Despite measuring light levels in the daylit space that are about 10 times higher than the light levels measured in the simulated classroom (light from screen projections augmented with multi-channel color tunable room lighting), after adapting to the environment, that large difference does not seem to be visually apparent to the participants.

Data collection is currently in process, detailed analysis of the human subjects data is ongoing, and the framework for comparisons of responses in the two environments is established. The results will be available in early summer for publication and presentation.

Recent advances in daylight research have generated increasingly sophisticated methods for measuring and evaluating light, including visual comfort indicators, environmental performance metrics, and non-visual lighting parameters. While these tools provide valuable information about luminous environments, a fundamental challenge remains: how can quantitative knowledge about light be translated into meaningful spatial understanding and design action?

This question is particularly relevant in architectural education, where students are often introduced to daylight through calculations, simulations, and performance indicators, yet may struggle to connect these data with the lived and sensory experience of space. In this context, learning to design with daylight requires not only the ability to measure light, but also the ability to perceive, interpret, and critically reflect upon its spatial, temporal, and experiential qualities.

This paper reflects on five international NLITED Summer Schools conducted between 2022 and 2025 as experimental learning environments for daylight design. Each edition addressed a different design challenge—including daylight simulation, adaptive reuse, daylight assessment, high-rise typologies, and historic building transformation—while combining environmental analysis, simulation tools, site observations, and design experimentation. Using a comparative qualitative approach, the study examines recurring learning and design patterns across the five editions. Rather than focusing on educational outcomes in a conventional sense, the analysis investigates how participants developed the capacity to move between measurable daylight parameters and experiential interpretations of light. Particular attention is given to the role of direct observation, spatial exploration, temporal awareness, interdisciplinary dialogue, and the critical use of simulation outputs within the design process.

The comparison reveals several recurring lessons. First, experiential engagement with space helps students contextualise and question quantitative daylight metrics. Second, temporal variations in daylight conditions often become meaningful only when they are directly observed and discussed. Third, interdisciplinary learning environments support a broader understanding of light as simultaneously measurable, perceptual, biological, and cultural. Finally, design decisions become more nuanced when sensory and experiential forms of knowledge complement objective performance data.

The paper argues that experimental learning environments can serve as translational settings where participants learn to connect measured daylight conditions with sensory, spatial, and experiential understanding. In doing so, they provide valuable insights for developing more human-centred approaches to daylight education and design, contributing to current discussions on how sensing light can bridge scientific knowledge and lived experience.