Anton Savov

Last Name

Savov

First Name

Anton

ORCID

0000-0002-5244-5285

Organisational unit

09566 - Dillenburger, Benjamin / Dillenburger, Benjamin

Show all metadata

Search Results

Publications 1 - 9 of 9

A Framework for the Design and Evaluation of Architectural Tilesets
Zhao, Hanbing; Savov, Anton; Zhang, Hang; et al. (2023)
eCAADe ~ Digital Design Reconsidered - Proceedings of the 41st Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2023) - Volume 2
Generative design, increasingly prevalent in architecture, enables design exploration and enhanced productivity compared to traditional methods. Researchers have investigated combinatorial design using tilesets, which encode architectural meaning and promote user-friendly interactions. However, most research focuses on discovering designs rather than fine-tuning tilesets. We propose a tile-based method that introduces metrics for evaluating generated layouts and tileset design space, addressing the research gap and facilitating practical applications. The design space evaluation feedback aids architects in customizing tilesets according to their objectives by exploring the impact of tile topology and rule changes. Our framework, illustrated through double-floor single-family house tilesets using the Wave Function Collapse algorithm, generates 3D designs and 2D layouts, enables minimal-specification diverse tilesets, and demonstrates fine-tuning to avoid grid-like monotonicity, a common limitation of tile-based generative design methods.
Identifying the Design Feature That Causes Project Delay in DfMA: A Dominant Element Analysis Method for Project Scheduling
Cao, Jianpeng; Zhang, Hang; Pan, Bo; et al. (2024)
Computing in Civil Engineering 2023: Visualization, Information Modeling, and Simulation
Design for manufacturing and assembly (DfMA) is an engineering methodology which aims to increase ease of manufacture and efficiency of assembly by considering manufacturing and assembly constraints in the design process. However, current DfMA approaches in the construction sector are not automated enough to identify the design features that may cause project delay in real time. This leads to longer design cycle. Also, current scheduling algorithms rely on human intervention to generate activity network from a design output. Addressing these inefficiencies, we propose an interpretative machining learning model to predict the construction duration given a design output. More importantly, the same model identifies the design features that may cause the most delay in the project. The model is trained on a residential design dataset with various features, such as layout, geometry, and element typology. The output of the model is the project duration and an importance map, indicating the influence each feature of the given design has on the total project duration. The results from this model can considerably reduce the design cycle by supporting architects to create fabrication and assembly aware design even when they have little knowledge of production and assembly processes. This model will contribute to a novel computational approach for DfMA.
MaskPLAN: Masked Generative Layout Planning from Partial Input
Zhang, Hang; Savov, Anton; Dillenburger, Benjamin (2024)
2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Layout planning, spanning from architecture to interior design, is a slow, iterative exploration of ill-defined problems, adopting a "I'll know it when I see it" approach to potential solutions. Recent advances in generative models promise automating layout generation, yet often overlook the crucial role of user-guided iteration, cannot generate full solutions from incomplete design ideas, and do not learn for the inter-dependency of layout attributes. To address these limitations, we propose Mask-PLAN, a novel generative model based on Graph-structured Dynamic Masked Autoencoders (GDMAE) featuring five transformers generating a blend of graph-based and image-based layout attributes. MaskPLAN lets users generate and adjust layouts with partial attribute definitions, create alternatives for preferences, and practice new composition-driven or functionality-driven workflows. Through cross-attribute learning and the user input as a global conditional prior we ensure that design synthesis is calibrated at every intermediate stage, maintaining its feasibility and practicality. Extensive evaluations show MaskPLAN's superior performance over existing methods across multiple metrics.
Energy-aware design: predicting building performance from layout graphs
Cao, Jianpeng; Zhang, Hang; Savov, Anton; et al. (2022)
Computing in Construction ~ Proceedings of the 2022 European Conference on Computing in Construction
Graph Neural Networks (GNNs) have become a popular toolkit for generative floor plan design. Although design variation has improved greatly, few studies consider non- geometrical characteristics, such as building energy performance, in the generative design process. This paper presents a GNN-based approach to predict the energy performance for floor plan customization (energy-aware design). The approach lays the foundation for a performance-aware generative design using GNN. The results show that the GNN can achieve high accuracy in energy performance prediction.
Real-Time Immersive Design with GenAI
Yang, Wenqian; Savov, Anton (2025)
At the Immersive Design Lab (IDL) in Design++ at ETH Zurich, we explore how generative AI can enrich architectural design and real-time interactive experiences. Our work leverages Stable Diffusion, CLIP, ControlNET, and ComfyUI to generate compelling, context-aware images that complement an immersive environment equipped with motion tracking (OptiTrack) and a 72-speaker audio system. The video attached to this submission showcases the interactive AI image generation aspect of the work.
Constructing Building Layouts and Mass Models with Hand Gestures in Multiple Mixed Reality Modes
Savov, Anton; Kessler, Martina; Reichardt, Lea; et al. (2023)
Towards Radical Regeneration
This paper presents a Mixed Reality framework for schematically defining building layouts and massing in multiple representations. Non-experts can use the framework to explore possible building configurations alone or in tandem with an architect. Our framework relies on a single-truth voxel matrix to track design changes and construct view-specific representations using the Marching Cubes and Marching Squares algorithms. We use only hand gestures for all design interactions instead of tangible objects or markers, to increase the mobility of users and make the application more accessible. The framework is tested in two prototypes for the HoloLens. The two prototypes have an objective to implement and test a variety of gestures for adding and removing volume, respectively area, from the designed building. The unified model representation across multiple MR views and interaction modes is the main contribution of this work and can be a valuable reference for the community developing applications of Mixed Reality in architecture. Additionally, we present a catalog of gesture-based interactions with the findings from our development process and the feedback from user studies.
Graph-Based Evolutionary Search for Optimal Hybrid Modularization of Building Construction Projects
Cao, Jianpeng; Said, Hisham; Savov, Anton; et al. (2024)
Journal of Construction Engineering and Management
Off-site construction has been a crucial part of industrializing the industry to realize higher productivity, better quality, and a more sustainable approach for constructing buildings. Off-site construction requires decomposing a floor plan into modules that can be in the form of either panelized walls or volumetric modules. However, the previous modularization models and approaches are limited due to their inability to consider the topological constraints of the modules, the flexible modularization of varying floor plans, and the mixed use of panelized walls and volumetric modules. As such, this paper proposes a graph-based optimization methodology for the hybrid modularization of building floor plans. The methodology was implemented using a multiobjective genetic algorithm that encodes and decodes the floor plan using novel graph modeling and operations. A visual programming script was developed to extract the wall properties, their adjacencies, and junction information from the building information model (BIM) of the floor plan. Time and cost estimation functions were developed to evaluate the hybrid strategies of panelized-volumetric modularization. The deployment of the methodology was demonstrated using an example floor plan design, which resulted in a spectrum of hybrid modularization plans ranging between fully volumetric and fully panelized solutions. For this specific example, the fully volumetric solution was 23% faster than the fully panelized solution but was 22% more expensive. The main contributions of this study are the topological modeling of module types, their floor plan postdesign flexible utilization, and the ability to explore hybrid modularization strategies. The findings of this study can prove useful for modular and off-site building manufacturers to improve their agility and increase their market share.
Generalist Generative Agent: Open-ended design exploration with large language models
Savov, Anton; Yoo, Angela; Lin, Che Wei; et al. (2025)
Architectural Informatics: Proceedings of the 30th International Conference of the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA 2025), Volume 2
Architects often navigate ambiguity in early-stage design by using metaphors and conceptual models to transform abstract ideas into architectural forms. However, current computational tools struggle with such exploratory processes due to narrowly defined design spaces. This paper investigates whether Large Language Models (LLMs) can offer an alternative generative paradigm by interpreting human intent and translating it into actionable design logic. We propose an Agentic AI framework in which LLM agents interpret metaphors, formulate design tasks, and generate procedural 3D models. Using this framework, we produced 1,000 procedural designs and 4,000 images based on 20 metaphors to demonstrate the emergent capabilities of LLMs for creating architecturally relevant conceptual models. Our findings suggest that LLMs effectively engage with ambiguity, delivering diverse, meaningful outputs with notable potential for early phase design. We discuss the strengths and shortcomings of the AI agents within the framework and suggest ways to extend their capacity for tackling open-ended design challenges, thereby enhancing their relevance in architectural practice.
DeepCanvas
Wu, Jiaqian; Bernhard, Mathias; Li, Li; et al. (2025)
Frontiers of Architectural Research
Building footprint synthesis in architectural design is a complex task that involves constraint satisfaction and objective optimization, traditionally relying on human expertise. This paper presents DeepCanvas, a novel framework that integrates procedural modeling with vision-based deep reinforcement learning (deep RL) to generate context-aware building footprints. Our approach represents designs through sequential parametric manipulations on primitive shapes, including geometric transformations and Boolean operations. The framework combines neural networks' pattern recognition capabilities with reward-driven exploration, enabling the RL agent to optimize architectural criteria through direct interaction with the design space rather than from existing samples. Although currently experimented with basic spatial quality metrics and simplified primitives for computational efficiency, results demonstrate that our agent successfully discovers context-adaptive strategies for generating optimized layouts from a vast design space while maintaining constructive outputs. The framework bridges machine-learnable and human-manipulable representations in architectural design, offering a systematic approach to learning-based design optimization that preserves procedural modeling’s interpretability and editability.

Publications 1 - 9 of 9

Anton Savov

Last Name

First Name

ORCID

Organisational unit

Filters

Search Results