kaopanboonyuen/KAO-DIFFSAT-SD-BASE-v1

🛰️ KAO-DIFFSAT-SD-BASE-v1

Clean Stable Diffusion Inpainting Backbone for Remote Sensing Research

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

🛰️ Visual Overview

<p align="center"> <img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reshow01.png" width="100%"> <img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reshow02.png" width="100%"> </p>

<p align="center"> <b>Satellite-oriented inpainting baseline enabling structured reconstruction across diverse geospatial scenes.</b> </p>

🌍 Overview

KAO-DIFFSAT-SD-BASE-v1 is a cleaned and structured Stable Diffusion inpainting backbone, prepared for remote sensing adaptation and research.

This repository is designed to serve as:

• —🧩 A baseline model for satellite image inpainting
• —🧠 A foundation for fine-tuning (KAO-DiffSat)
• —🔬 A reproducible starting point for geospatial generative AI

⚠️ Important: This repository does NOT introduce new training or optimization methods. It is a clean, minimal, and research-ready version of the original inpainting model.

🧠 Positioning

🏗️ Base Model

This model is derived from:

• —https://huggingface.co/sd2-community/stable-diffusion-2-inpainting

Original model developed by Stability AI and contributors.

🔹 What we did

• —Removed redundant files and duplicates
• —Organized model structure for clarity
• —Ensured compatibility with Diffusers
• —Prepared for downstream fine-tuning

👉 No modification to core weights has been introduced in this release.

🚀 Key Features

• —🧼 Clean and minimal model structure
• —⚡ Reduced storage footprint (no redundant weights)
• —🔁 Fully compatible with Diffusers pipelines
• —🛰️ Ready for remote sensing fine-tuning
• —🧩 Modular design for research experiments

🔬 Results Overview

<p align="center"> <img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reshow01.png" width="90%"> </p>

Qualitative reconstruction results using diffusion-based inpainting (baseline).

<p align="center"> <img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reshow02.png" width="90%"> </p>

Demonstration of spatial consistency and texture recovery.

🛰️ Full Visual Examples

<table> <tr> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall01.png"></td> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall02.png"></td> </tr> <tr> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall03.png"></td> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall04.png"></td> </tr> <tr> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall05.png"></td> <td><img src="https://huggingface.co/kaopanboonyuen/KAO-DIFFSAT-VHR-v1/resolve/main/img/reall06.png"></td> </tr> </table>

<p align="center"> <b>Example outputs across multiple scene types.</b> </p>

⚠️ What This Model Is (and Is Not)

✅ This model:

• —A clean baseline diffusion model
• —A directly usable inpainting backbone
• —A starting point for fine-tuning and research

❌ This model is NOT:

• —A fine-tuned satellite model
• —The KAO method
• —An optimized or domain-adapted system

👉 Those contributions will be introduced in future releases.

🚀 Usage (Diffusers)

from diffusers import StableDiffusionInpaintPipeline
import torch

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

prompt = "High-resolution satellite image with missing regions reconstructed realistically"

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

🛰️ Prompt Guide (Satellite-Oriented Usage)

To simulate remote sensing behavior, use prompts emphasizing:

• —spatial consistency
• —structural realism
• —geospatial coherence

🔹 Example

High-resolution satellite image of urban area,
reconstruct missing regions with realistic buildings,
maintain spatial consistency and natural textures

🧪 Use Cases

• —Cloud removal (baseline)
• —Image completion
• —Pre-training for remote sensing models
• —Generative data augmentation
• —Benchmark comparison

⚠️ Limitations

• —Prompt-dependent outputs
• —Not trained on real satellite datasets
• —No geospatial constraints
• —High GPU memory usage

🔮 Roadmap

• —🔜 KAO-DIFFSAT-INPAINTING-v1 (fine-tuned model)
• —🔜 KAO-integrated optimization
• —🔜 Prompt-free inference system

📚 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}
}

🔬 Project Context

This repository is part of the broader KAO-DiffSat research direction, aiming to bridge:

Diffusion Models × Remote Sensing × Optimization

🌐 Project Page

https://kaopanboonyuen.github.io/KAO/

🤝 Acknowledgements

• —Stability AI for foundational diffusion models
• —Hugging Face for open model hosting
• —Open-source community for tools and infrastructure

🌟 Vision

This work aims to advance:

Generative AI for Earth Observation

toward scalable and research-driven geospatial intelligence systems.