EarthSynth-180K is a multi-task, conditional, diffusion-based generative dataset designed for remote sensing image synthesis and understanding.
It was introduced in the paper "EarthSynth: Generating Informative Earth Observation with Diffusion Models" (arXiv 2025).

This dataset supports text-to-image generation, mask-conditioned synthesis, and multi-category augmentation for Earth observation research.

Dataset Details

Dataset Sources

Repository: GitHub - EarthSynth
Paper: ArXiv 2505.12108
Project Page: EarthSynth Website
Dataset Download: HuggingFace

Dataset Structure

Subset	# Images	Annotations	Format	Condition Types
Train	180,000	Masks, Prompts	PNG + JSONL	Mask + Text
Validation	10,000	Masks, Prompts	PNG + JSONL	Mask + Text
Augmented	180,000	Single-Category	PNG + JSONL	Category + Mask + Text

Masks: Binary/instance masks for each object category.
Prompts: Text prompts for conditional generation.
Augmentation: Single-category augmentation for CF-Comp training strategy.

Quick Start

from datasets import load_dataset

# Load dataset
dataset = load_dataset("jaychempan/EarthSynth-180K", split="train")

# Access one example
example = dataset[0]
print(example.keys())  # ['image', 'mask', 'prompt']

# Display image
from PIL import Image
import io

img = Image.open(io.BytesIO(example["image"]))
img.show()