rpradana36/genai_sustainablestreetscape

GenAI Sustainable Streetscape Dataset

Official dataset accompanying the manuscript

Reframing Urban "Sustainable" Streetscapes: Evidence from Cross-City Transformation and Convergence

!UNDER REVIEW!

Overview

This dataset contains paired Google Street View images, AI-generated sustainable streetscape transformations, and semantic segmentation outputs used to investigate how a zero-shot foundation model (OpenAI GPT-5 image-to-image) transforms urban streetscapes across different cities. The associated study evaluates whether AI-generated "sustainable" streets remain context-sensitive or converge toward a common visual prototype.

The dataset includes three cities representing distinct urban contexts:

City	Images
Jakarta	567
Melbourne	832
Singapore	694
Total	2,093

For every street-view image, the repository provides:

• Original Street View Image
• AI-generated baseline transformation
• Semantic segmentation outputs
• Pixel statistics
• Overlay visualization
• Panel visualization
• Alternative prompt generations
• Manual IoU annotations for segmentation validation

---

Dataset Structure

genai_sustainablestreetscape/
│
├── jakarta/
├── melbourne/
├── singapore/
│
├── alternative-prompts/
│
└── iou/

---

City folders

Each city folder contains the same directory structure.

jakarta/
│
├── image/
│      jakarta_2.png
│      ...
│
├── baseline/
│      jakarta_2_noncontext.png
│      ...
│
├── baseline-resize/
│      jakarta_2_noncontext.png
│      ...
│
├── point/
│      jakarta_points_streetview.csv
│
├── segmented-image/
│      overlay/
│      segmented/
│      panels/
│      pixel_counts.csv
│      pixel_counts.xlsx
│
├── segmented-baseline/
│      overlay/
│      segmented/
│      panels/
│      pixel_counts.csv
│      pixel_counts.xlsx
│
└── segmented-baseline-resize/
       overlay/
       segmented/
       panels/
       pixel_counts.csv
       pixel_counts.xlsx

The same directory structure is used for:

• melbourne/
• singapore/

---

Alternative Prompt Experiments

The alternative-prompts/ directory contains GPT-5 image-to-image generations produced using four prompting strategies. These experiments were designed to evaluate how different prompt formulations influence the visual interpretation of urban sustainability, and whether GPT-5 consistently converges toward similar streetscape transformations despite changes in prompt wording.

alternative-prompts/
├── baseline_resize/                (n = 208)
├── explicit_resize/                (n = 120)
├── non_vegetation_resize/          (n = 120)
└── sustainable_movement_resize/    (n = 120)

Each subfolder contains AI-generated streetscape images corresponding to one prompting strategy. The image naming convention follows the format:

jakarta_26_noncontext.png
jakarta_26_explicit.png
jakarta_26_non_vegetation.png
jakarta_26_sustainable_movement.png

Prompting Strategies

1. Baseline (Main Experiment)

The baseline prompt was used to generate all images analyzed throughout the main manuscript and served as the reference condition.

*Generate a realistic transformation of the attached street-view image (1024 × 1024) with the goal of making the streetscape appear more sustainable. Maintain the original structure and urban layout, but creatively reinterpret the scene through a sustainability-focused lens. Emphasize visual realism and plausible urban design improvements. The image is a street-view image taken in {city}, capturing its typical urban character.*

---

2. Explicit Multidimensional Sustainability

This prompt explicitly defines sustainability as a multidimensional concept, encouraging the model to consider multiple urban design interventions rather than inferring sustainability implicitly.

*Generate a realistic transformation of the attached street-view image (1024 × 1024) with the goal of making the streetscape appear more sustainable. Sustainability improvements may include, but are not limited to, pedestrian walkability, cycling infrastructure, public transit visibility, vegetation and greenery, and public social spaces. Maintain the original structure and urban layout while creatively reinterpreting the scene through a sustainability-focused lens. Emphasize visual realism and plausible urban design improvements. The image is a street-view image taken in {city}, capturing its typical urban character.*

---

3. Non-Vegetation Constraint

This prompt examines whether GPT-5 can generate sustainable streetscapes without relying on additional greenery, instead emphasizing transportation-related interventions.

*Generate a realistic transformation of the attached street-view image (1024 × 1024) with the goal of making the streetscape appear more sustainable, focusing specifically on pedestrian infrastructure, cycling infrastructure, and public transit elements. Do not add or expand vegetation, greenery, or natural landscaping. Maintain the original structure and urban layout while improving walkability, bikeability, and shared mobility infrastructure. Emphasize visual realism and plausible urban design improvements. The image is a street-view image taken in {city}, capturing its typical urban character.*

---

4. Sustainable Movement

This prompt prioritizes active mobility by emphasizing infrastructure supporting walking and cycling.

*Generate a realistic transformation of the attached street-view image (1024 × 1024) with the goal of making the streetscape more supportive of sustainable movement, combining pedestrian-friendly and cycling-friendly design. Emphasize wider and safer sidewalks, protected bicycle lanes, bicycle parking, and clearer pedestrian crossings while maintaining the original structure and urban layout. Emphasize visual realism and plausible urban design improvements. The image is a street-view image taken in {city}, capturing its typical urban character.*

Purpose

The alternative prompt experiments were conducted as a robustness analysis to evaluate the sensitivity of GPT-5's streetscape transformations to prompt wording. By systematically varying the conceptual emphasis—from a generic sustainability prompt to explicit multidimensional, vegetation-constrained, and mobility-focused prompts—the experiments investigate whether AI-generated sustainable streetscapes converge toward a common visual representation or remain responsive to different interpretations of sustainability.

---

IoU Validation

The iou/ folder contains manually annotated LabelMe polygons used to validate semantic segmentation.

Example:

jakarta_26_noncontext_labelme.json
melbourne_1_noncontext_labelme.json
singapore_10_noncontext_labelme.json

Approximately 10% of generated images from each city were manually annotated to compute Intersection over Union (IoU).

---

Semantic Categories

Mask2Former (Mapillary Vistas) predictions are regrouped into eight semantic categories:

• Sky
• Vegetation
• Built Structure
• Road Infrastructure
• Pedestrian Infrastructure
• Bikeability Infrastructure
• Vehicle
• Street Furniture

Pixel proportions for each image are provided in:

pixel_counts.csv
pixel_counts.xlsx

---

Image Naming Convention

Original image

jakarta_2.png

Baseline GPT-5 generation

jakarta_2_noncontext.png

Semantic segmentation

jakarta_2_noncontext_seg.png

Visualization overlay

jakarta_2_noncontext_overlay.png

Visualization panel

jakarta_2_noncontext_panel.png

---

Methodology

The complete workflow consists of five stages:

1. Street View image collection
2. GPT-5 image-to-image transformation
3. Semantic segmentation using Mask2Former
4. UMAP embedding and K-Means clustering
5. Homogenization and cross-city analysis

Further methodological details are available in the accompanying manuscript. :contentReference[oaicite:4]{index=4}

---

Applications

This dataset can be used for:

• Street-view semantic segmentation
• Urban visual perception research
• Sustainable streetscape analysis
• Generative AI evaluation
• Prompt engineering
• Image-to-image translation benchmarking
• Cross-city visual comparison
• Urban morphology studies

---

Citation

If you use this dataset, please cite:

@article{pradana2026genai,
  title={Reframing Urban "Sustainable" Streetscapes: Evidence from Cross-City Transformation and Convergence},
  author={Pradana, Mohammad Raditia and Gamal, Ahmad and Aryal, Jagannath},
  year={2026},
  note={Manuscript under review}
}

---

License

This dataset is released under the CC BY 4.0 license.

Please ensure that the usage of Google Street View imagery complies with Google's Terms of Service.

---

Contact

Mohammad Raditia Pradana

SMART CITY
Department of Geography
Universitas Indonesia

Email: mohammad.raditia03@ui.ac.id
Website: https://aditpradana36.github.io/
GitHub: https://github.com/AditPradana36

---

Acknowledgements

This work was supported by the Indonesian Endowment Fund for Education (LPDP), Universitas Indonesia, and The University of Melbourne.