license: apache-2.0
base_model:
- Qwen/Qwen-Image-Edit-2511
base_model_relation: adapter
pipeline_tag: image-text-to-image
tags:
- remote_sensing
- satellite
datasets:
- satellite-image-deep-learning/SODA-A
language:
- en
SatEdit
SatEdit is a mask-conditioned satellite image editing framework for localized object addition and removal. Given a satellite image, a binary mask, and a text instruction, SatEdit edits the selected region while preserving the surrounding overhead scene.
Highlights
- Mask-conditioned satellite image editing for object-level addition and removal.
- Localized edits guided by an input image, binary mask, and text prompt.
- ComfyUI and Hugging Face Diffusers inference workflows.
- LoRA fine-tuning for high-resolution image editing.
Qualitative comparisons
The figures below show qualitative editing examples and comparisons across models.
Semantic Composition of Dataset
Usage
Use the <satedit> trigger in prompts when running the SatEdit LoRA. For example:
<satedit> Add dense residential into the scene inside the provided mask
The model expects a satellite RGB image, a binary edit mask, and a text instruction describing the desired edit.
Diffusers inference
python "Diffusers Inference/infer_diffusers.py" \
--model_path /path/to/Qwen-Image-Edit \
--lora_path /path/to/satedit_lora \
--image /path/to/input.jpg \
--mask /path/to/mask.png \
--prompt "Add dense residential into the scene inside the provided mask" \
--output outputs/example.png
The complete inference code, ComfyUI workflow, training configuration, and installation instructions are available in the GitHub repository.
Project page
Visit the SatEdit project page for additional visual examples and project information.
License
Please review the licenses of this model, its base model, and all dependencies before use or redistribution.
Citation
If you find SatEdit useful in your research, please cite:
@misc{rizwan2026satedit,
title={SatEdit: A Mask-Conditioned Satellite Image Editing Framework for Localized Object Addition and Removal},
author={Muhammad Talha Rizwan and Muhammad Ahmed Amer},
year={2026},
eprint={2607.29367},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.29367}
}