🔥🔥🔥 ZoomEarth

We released ZoomEarth🌍, a vision language model that is designed to solve visual reasoning and question answering tasks on ultra-high-resolution remote sensing imagery with active perception. Moreover, ZoomEarth can seamlessly integrate with downstream models for tasks such as cloud removal, denoising, segmentation, and image editing through simple tool interfaces, demonstrating strong extensibility.

🎊 News and Updates

• 2026.2.21 🎉🎉🎉 ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks is accepted by CVPR 2026 (Highlight)
• 2025.11.22 🎉 We release the code that supports faster inference with vLLM!
• 2025.11.18 🎉 ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks is now avilable on arXiv!
• 2025.11.15 🎉 ZoomEarth-3B is publicly available on huggingface🤗!
• 2025.11.15 🎉 LRS-GRO is publicly available on huggingface🤗!

🎞️ Demo Video

https://github.com/user-attachments/assets/429a5ca9-6778-4e53-b4bf-dea32310c5e3

🧠 Model

Our model, ZoomEarth, is built upon Qwen2.5-VL-3B, a powerful VLM that
It supports fine-grained reasoning, spatial context interpretation, and multi-level object understanding.

• Model weights: ZoomEarth-3B 🤗
• Training scripts: src/train/
• Evaluate scripts: src/eval/

🛰️ Dataset

LRS-GRO contains high-resolution satellite images annotated with:

• Multi-level question types (global, regional, object)
• Bounding boxes and spatial relations
• Reasoning-based and factual QAs

Split	#Images	#Questions	Avg. Resolution
SFT	88	1011	5000
RL	228	2500	5000
Test	908	9734	5000

Download:
LRS-GRO 🤗

⚙️ Installation

Step 1. Create a conda environment and activate it.

conda create -n zoom-earth python=3.10 -y
conda activate zoom-earth

Step 2. Install PyTorch (We use PyTorch 2.4.1 / CUDA 12.1)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Step 3. Install other depencencies

pip install -r requirements.txt

Step 4. Configure NLTK local corpora (for WordNet)

# (1) Download WordNet data to a local directory (optional if already exists)
python -m nltk.downloader wordnet -d ./nltk_data

# (2) In your code, add the following before importing WordNet
import nltk

local_corpora = "./nltk_data"
nltk.data.path.insert(0, local_corpora)

from nltk.corpus import wordnet as wn

and then replace local_corpora with actual path in src/eval/eval.py, src/train/RL/src/open-r1-multimodal/src/open_r1/custom/customized_funcs.py

🚀 Quick start

python src/demo.py

🚂 Train

To train ZoomEarth, first run bash ./run_scripts/train_sft.sh to start SFT training phase.

After that, run bash ./run_scripts/train_rl.sh to start RL training phase.

📋 Test

To evaluate model on LRS-GRO, first run bash ./run_scripts/infer.sh to generate inference file.

After that, run bash ./run_scripts/eval.sh to get detailed evaluation result.

Or infer with vLLM:

First install vLLM to your environment And then start vLLM services by:

VLLM_USE_MODELSCOPE=true vllm serve \
PATH_TO_ZOOM_EARTH_MODEL \
--served-model-name ZoomEarth \
--max_model_len 2048 \
--host 0.0.0.0 \
--port 8000

Finally run python ./src/eval/infer_vllm.py --exp_name zoomearth-infer, and after infer you will find your result in result/zoomearth-infer.jsonl

📬 Contact

If you have questions or would like to collaborate, please contact us at:
📧 liuruixun6343@gmail.com

📧 HappyBug@stu.xjtu.edu.cn

Citation

If you found our work usful, welcome to cite us:

@InProceedings{Liu_2026_CVPR,
    author    = {Liu, Ruixun and Fu, Bowen and Song, Jiayi and Li, Kaiyu and Li, Wanchen and Xue, Lanxuan and Qiao, Hui and Zhang, Weizhan and Meng, Deyu and Cao, Xiangyong},
    title     = {ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {34877-34888}
}

❤ Acknowledgments

Thanks to the images from FAIR1M, GLH-Bridge and STAR, the benchmarks: LRS-VQA, MME-RealWorld, XLRS-Bench and GeoLLaVA-8K and the VLM-R1 training framework code.

_{© 2025 ZoomEarth Project. Released under the Apache 2.0 License.}