AI Generates Gymnasium Floor Plans Comparable to Winning Bid Designs

A diffusion-based artificial intelligence framework has generated a field-level floor plan for the Beijing Jiaotong University Campus Gymnasium in Xiong’an that researchers say is comparable to the winning bid plan actually implemented for the building. The study, published in the journal Architectural Intelligence, describes a two-stage generative workflow in which Stable Diffusion produces candidate layouts, automated assessment modules filter and rank them, and architects make the final choices between generation steps. The work targets large and medium-sized gymnasiums, a class of public buildings whose field-level plans must coordinate competition areas, spectator zones, circulation systems and support facilities under strict design standards.

The research team, led by Zhenyu Wang and Fang Zheng with Yufan Meng and Chen Xu, built the framework around a practical design project conducted in the author’s studio. The case study gymnasium has a total construction area of 31,810 square metres across three floors above ground and one underground, 3,600 fixed spectator seats, and a rectangular internal and external boundary layout. It combines a competition hall, basketball hall, swimming hall and badminton hall within one large-scale public building, while also serving as an emergency shelter venue. According to the authors, the project complies with stringent seismic, green building and sports facility standards, making it a representative test case for the methodology.

The dataset underpinning the system came from 200 completed gymnasiums collected manually as exemplary built cases. After screening against criteria drawn from the Sports Architecture Design Standards JGJ31-2003, which require auditorium capacities between 3,000 and 10,000 seats, the team retained 149 plan samples. Selected cases needed relatively complex functional zoning involving at least five functions, complete and centralized layouts, and diverse outer contour boundaries such as ellipses, rectangles and sawtooth polygons. Low-quality data with ambiguous functional information was screened out. The raw plans were then aligned in proportion using the 400-metre runway track as a benchmark, and colour labelling followed the International Olympic Committee operational design standard to ensure consistency.

From this material the researchers produced three forms of data: block plans showing functional zones as simple colour blocks, detailed plans that add room divisions, and functional topology diagrams in which node size reflects the area ratio of each function. Computer vision techniques, multimodal models and large language models were used to extract edge adjacency relationships and calculate node sizes for the topology diagrams. Through data enhancement operations such as mirror flipping and normalization, the final training sets grew to 434 block plans, 434 detailed plans and 434 labelled topology diagrams. The first two sets trained the generative model, while the third trained the assessment model. The team notes that flipping constraints mattered: for outdoor gymnasiums where VIP seats must sit on the west side, horizontal flipping was disallowed.

In the first generation stage, Stable Diffusion version 1.5 was fine-tuned with Low-Rank Adaptation, or LoRA, and constrained by ControlNet, with the gymnasium boundary line fixed in advance by architects. Prompts emphasized the number, area and adjacency of functional zones. The LoRA model trained for 100 epochs with one model saved every 10 epochs, and the 80th-epoch model with the lowest loss, 0.01386, was used for generation. Sampling used the DPM++ 2M sampler over 35 steps with a refiner switching timing of 0.36, a CFG scale of 8 and a seed of -1. For the case study, the ControlNet boundary was a rectangle of 124 by 68 metres, and by varying LoRA training epochs and weights the team obtained 120 block plans in batch mode.

Assessment of those outputs proceeded in three sequential stages. Rule-based screening implemented through computer vision eliminated images of poor visual quality, minimized colour confusion and boundary overflow, required at least one Field of Play functional area carrying the colour code #7FB1D3, and verified that the aspect ratio of the Field of Play complied with JGJ31-2003. An adapted convolutional neural network then ranked the survivors by topological similarity to the exemplary built cases, operating not on raw floor plan images but on simplified topology diagrams in which colour and node size encode function and area. The network’s final classification layer applies a SoftMax function assigning mutually exclusive good or bad labels with discriminative scores, and the good-label probability serves as the topological similarity score. Trained with cross-entropy loss and an SGD optimizer at a learning rate of 0.001, the model stabilized after 20 epochs with training loss between 0.05 and 0.18 and test accuracy consistently above 0.99.

Architects then reviewed the ten highest-ranked block plans and selected one for the second stage. That chosen block plan entered ControlNet as a geometric constraint for image-to-image generation, so the second-stage outputs were refined from the first-stage result rather than generated from scratch. A second LoRA model, trained on the detailed plan dataset for 70 epochs to a minimum loss of 0.0146, produced 30 detailed plans that subdivided the larger functional zones into room units and circulation networks. The second assessment stage applied a two-level system. Project-specific metrics, extracted from the design brief, checked requirements such as competition halls and basketball courts on the ground floor, an eastern comprehensive competition hall, a western basketball and shared hall, offices along the southern and eastern sides, entrance lobbies on both sides of the basketball hall, a dedicated field entrance on the eastern side of the competition hall, and electrical switching and transformer substations on the northern side.

The general metric measured functional colour consistency between each detailed plan and the input block plan using the perceptual colour difference index CIEDE2000. Because the detailed plans introduce new wall-line strokes that are high-contrast and near-neutral, a morphology- and gradient-based preprocessing step detected and masked those pixels in both images, excluding them from the colour-difference computation. Lower ΔE values indicate that the functional colour regions of the generated plan remain more consistent with the input design. The 30 candidates were ranked in ascending order and the top 12 presented to the architect, who selected the first scheme as the final gymnasium plan based on professional experience. In that plan, the two strip-shaped spaces of the Venue Operations area were maintained while their interiors were subdivided with room walls and corridors introduced along the boundaries.

The comparison with the implemented winning bid plan showed strong alignment. Both schemes locate the Field of Play centrally as the organizational core, position the Athlete Training Area on the left with a coherent athlete circulation system and dedicated service areas on both sides, keep the Spectator/Visitor Area adjacent to the field for direct visibility, cluster the Press Area on the right for event operations, and place the Commercial District near the entrance to maximize visitor access and commercial value. The generated detailed plans with room separations also showed close consistency: the implemented plan maintains two strip-shaped spaces in the Venue Operations area, and the generated plan preserves that logic while introducing room units and boundary corridors. The authors state that these results demonstrate the framework can reproduce essential organizational principles typically ensured by professional architects, achieving outcomes comparable to winning bid schemes.

The filtering pipeline quantitatively compressed the human decision burden. From 120 generated block plans, rule-based screening and CNN-based topological similarity evaluation reduced the pool to 10 high-quality candidates, a reduction of over 90 percent in manual screening workload. In the second stage, 30 detailed plans were filtered by the project-specific and general metrics down to the top 12 for final inspection, from which a single optimal scheme was selected. The final choice was supported by two independent quantitative indicators: the selected block plan ranked among the top group in CNN-based topological similarity score, and the corresponding detailed plan passed the project-specific screening and exhibited one of the lowest masked CIEDE2000 values in its batch. The authors frame the contribution as a feedback-enabled generative-assessment paradigm that embeds explicit evaluation modules between generation steps, maintaining consistency from functional zoning to room-level layouts. They also acknowledge limitations: the gymnasium plan dataset remains limited, the study relies on Stable Diffusion v1.5 rather than newer models, a CNN was used instead of graph neural networks for consistency with the visual-spatial representation, quantitative validation covered only this single representative project, and performance criteria such as circulation efficiency and sightline quality, which require 3D spatial data, crowd-flow simulation and visibility analysis, have not yet been incorporated.

Subject of Research: Diffusion-based deep learning generation and assessment of gymnasium field-level floor plans

Article Title: Intelligence assessment for diffusion-based generation of gymnasium field-level plans using deep learning

Article References: Wang, Z., Meng, Y., Xu, C., & Zheng, F. (2026). Intelligence assessment for diffusion-based generation of gymnasium field-level plans using deep learning. Architectural Intelligence, 5(1), Article 6. https://doi.org/10.1007/s44223-026-00114-w

Image Credits: AI Generated

DOI: 10.1007/s44223-026-00114-w

Keywords: generative AI, Stable Diffusion, LoRA, ControlNet, convolutional neural network, gymnasium design, floor plan generation, architectural intelligence, CIEDE2000, functional zoning, topological similarity, Beijing Jiaotong University