Datasets:

jp-cypress
/

Procedural-City-Multimodal-Dataset

license: cc-by-nc-4.0
task_categories:
  - image-to-image
  - text-to-image
tags:
  - synthetic
  - blender
  - urban-planning
  - controlnet
  - multimodal
size_categories:
  - 10K<n<100K

Synthetic Urban Multimodal Dataset (v0.1)

This dataset contains 8,000 high-quality synthetic images of procedurally generated city models, designed specifically for training and evaluating AI models such as ControlNet and LoRA.

Dataset Description

All assets were generated using a custom Blender-Python automated pipeline. The dataset provides 100 unique urban architectures, each rendered from 8 directions with 10 different artistic styles.

Total Samples: 8,000 (Total 40,000 files across modalities)
Resolution: 512x512
Modalities per sample:
- rgb_images: Stylized urban rendering.
- depth_maps: High-precision depth information.
- normal_maps: Surface normal vectors.
- albedo_images: Pure texture color (unlit).
- mask_images: Binary/Alpha masks for segmentation.

Data Structure

The dataset follows the standard Hugging Face metadata format. Each entry in metadata.jsonl maps the RGB image to its corresponding multimodal maps and captions.

Intended Use

Training ControlNet for spatial and structural control.
Fine-tuning LoRA for specific architectural or artistic styles.
Research in Computer Vision and Synthetic Data generation.

License

This dataset is provided under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Commercial use is prohibited.

Author & Technical Background

Developed by [ひのき(jp-cypress) / https://zenn.dev/jp_cypress]. For technical details on how this dataset was procedurally generated using Blender API and Python, please refer to the technical report on Zenn (Japanese).