Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CoMa: Contextual Massing Generation with Vision-Language Models

Evgenii Maslov, Valentin Khrulkov, Anastasia Volkova, Anton Gusarov, Andrey Kuznetsov, Ivan Oseledets

TL;DR

The paper tackles the challenge of automated building massing generation by formulating it as a conditional generation task conditioned on functional requirements, site contours, and urban context. It introduces CoMa-20K, a multi-modal dataset that pairs 3D massing geometries with programmatic data and visual context, and presents a Vision-Language Model benchmark evaluating both fine-tuned and zero-shot models. Key findings show that while large fine-tuned models can generate complex geometries, they often produce artifacts, whereas large zero-shot models yield clean but overly simple massings with limited contextual adaptation; overall, there remains a gap in robust geometric reasoning and context integration. The work establishes a foundational benchmark and dataset for data-driven architectural design and highlights avenues for future research in specialized geometric reasoning and balanced multi-modal datasets, aiming to enable AI-assisted design tools that complement human creativity.

Abstract

The conceptual design phase in architecture and urban planning, particularly building massing, is complex and heavily reliant on designer intuition and manual effort. To address this, we propose an automated framework for generating building massing based on functional requirements and site context. A primary obstacle to such data-driven methods has been the lack of suitable datasets. Consequently, we introduce the CoMa-20K dataset, a comprehensive collection that includes detailed massing geometries, associated economical and programmatic data, and visual representations of the development site within its existing urban context. We benchmark this dataset by formulating massing generation as a conditional task for Vision-Language Models (VLMs), evaluating both fine-tuned and large zero-shot models. Our experiments reveal the inherent complexity of the task while demonstrating the potential of VLMs to produce context-sensitive massing options. The dataset and analysis establish a foundational benchmark and highlight significant opportunities for future research in data-driven architectural design.

CoMa: Contextual Massing Generation with Vision-Language Models

TL;DR

The paper tackles the challenge of automated building massing generation by formulating it as a conditional generation task conditioned on functional requirements, site contours, and urban context. It introduces CoMa-20K, a multi-modal dataset that pairs 3D massing geometries with programmatic data and visual context, and presents a Vision-Language Model benchmark evaluating both fine-tuned and zero-shot models. Key findings show that while large fine-tuned models can generate complex geometries, they often produce artifacts, whereas large zero-shot models yield clean but overly simple massings with limited contextual adaptation; overall, there remains a gap in robust geometric reasoning and context integration. The work establishes a foundational benchmark and dataset for data-driven architectural design and highlights avenues for future research in specialized geometric reasoning and balanced multi-modal datasets, aiming to enable AI-assisted design tools that complement human creativity.

Abstract

The conceptual design phase in architecture and urban planning, particularly building massing, is complex and heavily reliant on designer intuition and manual effort. To address this, we propose an automated framework for generating building massing based on functional requirements and site context. A primary obstacle to such data-driven methods has been the lack of suitable datasets. Consequently, we introduce the CoMa-20K dataset, a comprehensive collection that includes detailed massing geometries, associated economical and programmatic data, and visual representations of the development site within its existing urban context. We benchmark this dataset by formulating massing generation as a conditional task for Vision-Language Models (VLMs), evaluating both fine-tuned and large zero-shot models. Our experiments reveal the inherent complexity of the task while demonstrating the potential of VLMs to produce context-sensitive massing options. The dataset and analysis establish a foundational benchmark and highlight significant opportunities for future research in data-driven architectural design.
Paper Structure (16 sections, 4 figures, 3 tables)

This paper contains 16 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Dataset
  4. Dataset Overview
  5. Data Schema and Properties
  6. Data Sources
  7. Dataset Collection Pipeline
  8. Statistics and Splitting
  9. Method
  10. Finetuned VLM Approach
  11. Zero-shot VLM Approach
  12. Experiments
  13. Implementation Details
  14. Quantitative Results
  15. Qualitative Results
  16. ...and 1 more sections

Figures (4)

  • Figure 1: A complete sample from the CoMa-20K dataset, illustrating the massing generation task. The input defines the problem through multiple modalities: the functional and economical requirements specify the building program, the site contour defines the legal boundaries for development, and the contextual views describe the urban environment. The ground truth output is the 3D massing geometry that satisfies these inputs. In this visualization, the generated massing (in gray) is shown integrated into its urban context (in gray-blue), demonstrating a plausible and context-aware design solution.
  • Figure 2: Dataset Creation Pipeline. The pipeline for constructing the CoMa-20K dataset consists of three main stages: Spatial Preprocessing, where raw geospatial data is converted to a metric system and prepared for merging; Massing Creation, where building geometries are assembled and fused with their functional-economical metadata; and Environment Creation, where the visual context for each site is generated by assembling neighboring buildings and producing multi-style renderings.
  • Figure 3: CoMa-20K Dataset Statistics. Distributions of key attributes across the dataset: (a) building functions, (b) number of buildings per site, (c) dwelling units per building, (d) commercial spaces per building, (e) total usable area, (f) number of floors, and (g) public capacity of commercial spaces.
  • Figure 4: Learning curves for the fine-tuned models.