Text2Earth: Unlocking Text-driven Remote Sensing Image Generation with a Global-Scale Dataset and a Foundation Model

**[Chenyang Liu](https://chen-yang-liu.github.io/), [Keyan Chen](https://kyanchen.github.io), [Rui Zhao](https://ruizhaocv.github.io/), [Zhengxia Zou](https://scholar.google.com.hk/citations?hl=en&user=DzwoyZsAAAAJ), and [Zhenwei Shi*✉](https://scholar.google.com.hk/citations?hl=en&user=kNhFWQIAAAAJ)** [](https://chen-yang-liu.github.io/Text2Earth/) [](https://ieeexplore.ieee.org/document/10988859) [](https://youtu.be/Rw9wzUpO01M)

Share us a :star: if you're interested in this repo

This is official repository of the paper: "Text2Earth: Unlocking Text-driven Remote Sensing Image Generation with a Global-Scale Dataset and a Foundation Model", accepted by IEEE Geoscience and Remote Sensing Magazine.

## News 🔥

✅ 2025.06.01: Git-RSCLIP series downloads exceeded 60,000 times 🔥

✅ 2025-04-16: The paper has been accepted by IEEE Geoscience and Remote Sensing Magazine.

✅ 2025-03-03: Our Git-RSCLIP model available: [🤗 Huggingface | 🌊 Modelscope]

✅ 2025-02-20: The Git-10M dataset is available: [🤗 Huggingface | 🌊 Modelscope].

✅ 2025-01-01: The paper is available.

Table of Contents

• 🛰️ Git-10M Dataset
  • Dataset Download
  • Visual Quality Enhancement
• 🧩 Text2Earth Model
  • Pre-trained Weights
  • Demo-Usage
  • Installation
  • Training
  • Evaluation
• 🍀 Git-RSCLIP Model
• ✍️ Citation

🛰️ Git-10M Dataset

Dataset Download

• The Git-10M dataset is a global-scale dataset, consisting of 10.5 million image-text pairs with geographical locations and resolution information.

• The Git-10M dataset is available at: [🤗 Huggingface | 🌊 Modelscope].

  
  

  
  
  

Visual Quality Enhancement

• You can skip the following steps if you have higher visual quality requirements for the image.

• Some collected images exhibited poor visual quality, such as noise and artifact, which could negatively impact the training of image generation models. To address this, you can use an image enhancement model pre-trained on my private high-quality remote sensing dataset to improving the overall image quality.

  Follow the steps below:

  Step 1: python cd ./Text2Earth/Tools Step 2: Run Python code to process images:

  python python visual_quality_enhancement.py \ --input_dir /path/to/Git-10M/images \ --output_dir /path/to/Git-10M/enhanced_images

🧩 Text2Earth Model

Pre-trained Weights

We provide two versions of the model: - Text2Earth Link : [🤗 Huggingface | 🌊 Modelscope]. - Text2Earth-inpainting Link : [🤗 Huggingface | 🌊 Modelscope].

Demo-Usage:

✅ Loading Usage 1: Use Text2Earth directly through [🤗 Diffuser] without installing our repository.

✅ Loading Usage 2: Install our repository (See Installation), then you can use the provided Pipeline, which is more convenient for users to customize and edit.

✅ NOTE: Text2Earth and Text2Earth-inpainting allow users to specify the spatial resolution of the generated images, ranging from 0.5m to 128m per pixel. This can be achieved by including specific identifiers in the prompt. ```python

You can indirectly set the spatial resolution by specifying the GoogleMapLevel, which ranges from [10, 18], corresponding to resolutions from [128m, 0.5m].

The conversion formula is: Resolution = 2^(17 - Level).

GoogleMapLevel = 16 # Resolution = 2**(17-Level) contentprompt = "Seven green circular farmlands are neatly arranged on the ground" promptwithresolution = '{res}GOOGLELEVEL' + contentprompt pipe = xxx # Text2EarthDiffusionPipeline or Text2EarthDiffusionInpaintPipeline image = pipe(prompt=promptwith_resolution, ...).images[0] ```

Installation

Training

Code is coming soon.

Evaluation

Code is coming soon.

Experimental Results

Building on the Git-10M dataset, we developed Text2Earth, a 1.3 billion parameter generative foundation model. Text2Earth excels in resolution-controllable text2image generation and demonstrates robust generalization and flexibility across multiple tasks.

• Comparison of Text2image models on the previous benchmark dataset (RSICD):

On the previous benchmark dataset RSICD, Text2Earth surpasses the previous models with a significant improvement of +26.23 FID and +20.95% Zero-shot OA metric.

• Zero-Shot text2image generation: Text2Earth can generate specific image content based on user-free text input, without scene-specific fine-tuning or retraining.
  

  
  
  
  

• Unbounded Remote Sensing Scene Construction: Using our Text2Earth, users can seamlessly and infinitely generate remote sensing images on a canvas, effectively overcoming the fixed-size limitations of traditional generative models. Text2Earth’s resolution controllability is the key to maintaining visual coherence across the generated scene during the expansion process.
  

  
  
  

• Remote Sensing Image Editing: Text2Earth can perform scene modifications based on user-provided text such as replacing or removing geographic features. And it ensures that these modifications are seamlessly integrated with the surrounding areas, maintaining continuity and coherence.
  

  
  
  

• Cross-Modal Image Generation: Text2Earth can be used for Text-Driven Multi-modal Image Generation, including RGB, SAR, NIR, and PAN images.
  

  
  
  

Text2Earth also exhibits potential in Image-to-Image Translation, containing cross-modal translation and image enhancement, such as PAN to RGB (PAN2RGB), NIR to RGB (NIR2RGB), PAN to NIR (PAN2NIR), super-resolution, and image dehazing.

🍀 Git-RSCLIP Model

• The Git-RSCLIP model is a remote sensing image-text foundation model, which is trained on the Git-10M dataset.
• For more details, please see the github repository: [Github]

✍️️ Citation

If you find this paper useful in your research, please consider citing: @ARTICLE{10988859, author={Liu, Chenyang and Chen, Keyan and Zhao, Rui and Zou, Zhengxia and Shi, Zhenwei}, journal={IEEE Geoscience and Remote Sensing Magazine}, title={Text2Earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model}, year={2025}, volume={}, number={}, pages={2-23}, doi={10.1109/MGRS.2025.3560455}}

📖 License

This repo is distributed under MIT License. The code can be used for academic purposes only.