jp-cypress/Procedural-City-Multimodal-Dataset-V2

Synthetic Urban Multimodal Dataset - Volume 2 (v0.2)

This is the second major release of the Synthetic Urban Multimodal Dataset, part of an ongoing observation project on AI-generated data ecosystems.

Volume 2 introduces a completely new collection of urban environments to the series, providing 40,000 high-precision files (8,000 completely new unique scenes × 5 modalities) generated entirely via the "Constructive Furnace" (our custom Blender-Python pipeline).

What's New in Volume 2

• —Brand New Urban Models: This release features 100 completely different city architectures and 8,000 new procedural seeds, entirely independent of Volume 1.
• —Consistent High Quality: Maintained the exact same robust pipeline. Each scene is strictly aligned across 5 different modalities, ensuring drop-in compatibility with Volume 1 for ControlNet, LoRA, and VLM training.
• —Massive Data Volume: Provides 40,000 completely new files, delivering the bulk data scale required for robust AI fine-tuning.

Dataset Description

The dataset provides 100 new unique urban architectures, each rendered from 8 cardinal directions with 10 distinct artistic styles (ranging from realistic blueprints to stylized anime).

• —Total Samples: 8,000 unique scenes
• —Total Files: 40,000 (RGB + 4 Maps)
• —Resolution: 512x512 (Lossless PNG)
• —Modalities per sample:
• —rgb_images: High-quality stylized rendering.
• —depth_maps: High-precision 16-bit-like depth information.
• —normal_maps: Geometric surface vectors for relighting.
• —albedo_images: Pure texture color without lighting/shadows.
• —mask_images: Semantic masks for object segmentation.

Technical Foundation: "Constructive Furnace"

This dataset is 100% Synthetic. No human-painted assets or copyrighted photos were used. The entire pipeline is built on Blender's Python API (bpy), ensuring a noise-free, mathematically perfect ground truth for AI training.

Data Structure

Each entry in metadata.jsonl follows the standard Hugging Face format for seamless integration with the datasets library.

{
  "file_name": "rgb_images/World_City_20260404_003156_anime_line_art_000.png",
  "text": "3d render, isometric, view from 0 degrees, anime and line art style, simple background, transparent background",
  "depth_file": "depth_maps/World_City_20260404_003156_anime_line_art_000.png",
  "normal_file": "normal_maps/World_City_20260404_003156_anime_line_art_000.png",
  "albedo_file": "albedo_images/World_City_20260404_003156_anime_line_art_000.png",
  "mask_file": "mask_images/World_City_20260404_003156_anime_line_art_000.png",
  "metadata": {
    "style": "anime",
    "angle": 0,
    "line_art": true,
    "model_name": "World_City"
  }
}