🛰️ KAO-DIFFSAT-VHR-v1 (Beta Version)

Diffusion-based Satellite Image Inpainting (VHR) — Baseline Release

🚀 Project website: https://kaopanboonyuen.github.io/KAO/

---

🛰️ Visual Overview

High-fidelity satellite image reconstruction across diverse scenes (urban, agricultural, coastal).

---

🌍 Overview

KAO-DIFFSAT-VHR-v1 is a satellite-domain adaptation of diffusion-based inpainting, designed for Very High Resolution (VHR) remote sensing imagery.

This release serves as a baseline model in our IEEE TGRS work:

KAO: Kernel-Adaptive Optimization in Diffusion for Satellite Image

🎉 Accepted to IEEE Transactions on Geoscience and Remote Sensing (TGRS)
🌏 Presentation at AOGS 2026, Japan

⚠️ Important:
This model is NOT the KAO method itself.
It is a Stable Diffusion–based inpainting backbone adapted for satellite imagery, used as one of the reference baselines in our study.

👉 The full KAO framework (training-free / prompt-free optimization) will be released separately.

---

🧠 Positioning

Component	Description
This Repo	Diffusion baseline for satellite inpainting
KAO (Paper)	Optimization-based enhancement (no retraining required)
Future Release	KAO-integrated inference (no prompt dependency)

---

🏗️ Base Model

This model is built upon:

• Stable Diffusion v2 Inpainting
• Provided via the Hugging Face ecosystem

We adapt the pipeline for remote sensing characteristics, including:

• Large spatial structures (roads, rivers, urban layouts)
• Texture consistency across wide regions
• Geometric coherence in aerial perspective

---

🚀 Key Features

• 🛰️ Satellite-aware inpainting behavior
• 🌆 Improved structural consistency for urban layouts
• 🌿 Better handling of vegetation and natural textures
• 🧩 Compatible with standard diffusion pipelines (Diffusers)

---

🔬 Results Overview

Qualitative comparison across multiple models. KAO demonstrates superior structural reconstruction and texture consistency.

Detailed comparison highlighting fine-grained structures and realistic spatial coherence.

---

🛰️ Full Benchmark Results

Comparison across multiple scenes including urban, agricultural, and occluded environments.
KAO consistently preserves spatial structure, enhances texture fidelity, and improves reconstruction realism.

---

⚠️ What This Model Is (and Is Not)

✅ This model:

• A baseline diffusion model for satellite inpainting
• Used for comparative evaluation in our research
• Suitable for prompt-based reconstruction tasks

❌ This model is NOT:

• The full KAO algorithm
• A training-free optimization method
• A prompt-free system

👉 Those capabilities are introduced in our paper and will be released in future versions.

---

🚀 Usage (Diffusers)

from diffusers import StableDiffusionInpaintPipeline
import torch

pipe = StableDiffusionInpaintPipeline.from_pretrained(
    "kaopanboonyuen/KAO-DIFFSAT-VHR-v1",
    torch_dtype=torch.float16
).to("cuda")

prompt = "High-resolution satellite image of urban area with missing regions reconstructed"

image = pipe(
    prompt=prompt,
    image=init_image,
    mask_image=mask
).images[0]

---

🛰️ Prompt Guide (Satellite Domain)

Use prompts that enforce:

• spatial coherence
• structural realism
• geospatial consistency

🔹 Template

High-resolution satellite image of [SCENE],
reconstruct missing regions with realistic [STRUCTURE],
maintain spatial consistency, natural textures, and accurate geometry

---

🧪 Use Cases

• Cloud removal in satellite imagery
• Infrastructure reconstruction
• Map completion
• Environmental monitoring
• Disaster recovery (flood, wildfire)

---

⚠️ Limitations

• Prompt-sensitive behavior
• Not optimized for hyperspectral imagery
• High memory footprint (~25GB)
• No explicit geospatial constraints (baseline only)

---

🔮 Roadmap

• 🔜 KAO-DIFFSAT-v2 (KAO-integrated)
• 🔜 Prompt-free inference (as proposed in paper)
• 🔜 Lightweight and scalable variants

---

📚 Citation

🔹 Base Model

@misc{stabilityai2022sd2,
  title        = {Stable Diffusion v2},
  author       = {{Stability AI}},
  year         = {2022},
  howpublished = {\url{https://github.com/Stability-AI/stablediffusion}},
  note         = {Including inpainting variant}
}

🔹 Our Work (KAO)

@article{panboonyuen2025kao,
  title={KAO: Kernel-Adaptive Optimization in Diffusion for Satellite Image},
  author={Panboonyuen, Teerapong},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2025},
  publisher={IEEE}
}

---

🔬 Project Overview

This repository presents KAO-DIFFSAT-VHR-v1, a diffusion-based framework for very high-resolution (VHR) satellite image inpainting, designed for geospatial restoration and reconstruction tasks.

The model builds upon Stable Diffusion-based inpainting and adapts it for remote sensing imagery.

---

📄 Paper

📘 Full paper available on arXiv:
https://arxiv.org/abs/2511.02462

---

🌐 Project Page

🚀 Project website:
https://kaopanboonyuen.github.io/KAO/

---

🤝 Acknowledgements

• Stability AI for foundational diffusion models
• Hugging Face for open model hosting
• Remote sensing community for datasets and benchmarks

---

🌟 Vision

This project aims to advance:

Generative AI × Earth Observation

toward scalable, reliable, and interpretable satellite intelligence systems.

---