MarsScope-v1

MarsScope-v1

MarsScope-v1 is a planetary vision-language model and aligns orbital imagery and scientific language in a shared embedding space, supporting text-to-image, image-to-image, and multimodal retrieval for label-free exploration of planetary surfaces. It is based on ViT-L-14-quickgelu, initialized from CLIP-DFN2B, and fine-tuned at 512 x 512 resolution on more than 200,000 curated pairs of planetary surface images and geomorphological descriptions.

Model Details

• Architecture: ViT-L-14-quickgelu
• Framework: OpenCLIP
• Initialization: CLIP-DFN2B
• Input resolution: 512 x 512
• Parameters: approximately 0.4B
• Training objective: contrastive image-text alignment

Applications

The MarScope framework was introduced in Natural Language-Driven Global Mapping of Martian Landforms for natural language-driven global mapping, process-oriented geomorphological retrieval, and visual search for rare or previously unmapped features.

It was subsequently evaluated in MarsRetrieval, which covers paired Martian image-text retrieval, fine-grained landform retrieval, and planetary-scale geo-localization.

Usage

For large-scale retrieval, encode and normalize the image and text queries separately, then rank images by cosine similarity. To illustrate the effect of planetary-domain fine-tuning, we compare the image-text cosine similarities produced by the CLIP-DFN2B base model and MarsScope-v1:

pip install torch open_clip_torch pillow huggingface_hub

import open_clip
import torch
from huggingface_hub import hf_hub_download
from PIL import Image

repo_id = "claytonwang/MarsScope-v1"
model_id = f"hf-hub:{repo_id}"
device = "cuda" if torch.cuda.is_available() else "cpu"

model, _, preprocess = open_clip.create_model_and_transforms(model_id)
tokenizer = open_clip.get_tokenizer(model_id)
model = model.to(device).eval()

image_names = [
    "central_peak_crater.png",
    "yardangs.png",
]
text_prompts = [
    "a satellite image of a central peak crater on Mars",
    "a satellite image of yardangs on Mars",
]

image_paths = [
    hf_hub_download(repo_id=repo_id, filename=name)
    for name in image_names
]
images = torch.stack([
    preprocess(Image.open(path).convert("RGB"))
    for path in image_paths
]).to(device)
text = tokenizer(text_prompts).to(device)

with torch.inference_mode():
    image_features = model.encode_image(images, normalize=True)
    text_features = model.encode_text(text, normalize=True)
    similarity = image_features @ text_features.T

# Rows: central_peak_crater.png, yardangs.png
# Columns: crater prompt, yardangs prompt
print(similarity)

Examples

CLIP-DFN2B

tensor([[0.3569, 0.2822],
        [0.2729, 0.2723]])

MarsScope-v1

tensor([[0.4516, 0.0585],
        [0.0963, 0.3537]])

Compared with the base CLIP-DFN2B, MarsScope-v1 shows stronger alignment between Martian landform images and their matching descriptions while more clearly separating mismatched pairs. For comprehensive evaluation, refer to MarsRetrieval.

Limitations

MarsScope-v1 is intended for planetary-science research. Retrieval results may be affected by image resolution, illumination, spatial scale, geographic coverage, and the terminology used in a query. Model outputs should be validated by domain experts before scientific interpretation.

Citation

@article{wang2026natural,
  title   = {Natural Language-Driven Global Mapping of Martian Landforms},
  author  = {Wang, Yiran and Wang, Shuoyuan and Wei, Zhaoran and Zhao, Jiannan
             and Yao, Zhonghua and Xie, Zejian and Zhang, Songxin and Huang, Jun
             and Jing, Bingyi and Wei, Hongxin},
  journal = {arXiv preprint arXiv:2601.15949},
  year    = {2026}
}

@article{wang2026marsretrieval,
  title   = {MarsRetrieval: Benchmarking Vision-Language Models for
             Planetary-Scale Geospatial Retrieval on Mars},
  author  = {Wang, Shuoyuan and Wang, Yiran and Wei, Hongxin},
  journal = {arXiv preprint arXiv:2602.13961},
  year    = {2026}
}