README.md

---
license: openrail++
tags:
- stable-diffusion
- diffusers
- image-to-image
- remote-sensing
library_name: diffusers
pipeline_tag: text-to-image
---

# Stable Diffusion 2.1 Base

This repository mirrors the original diffusers-format `stabilityai/stable-diffusion-2-1-base` checkpoint for reproducible public use.

⚠️ This repository is an unofficial mirror and is not affiliated with Stability AI.

The model weights and configuration files in this repository are provided under the original **CreativeML Open RAIL++-M** license. Please read [LICENSE-MODEL](./LICENSE-MODEL) before using the model.

## Original Source And Attribution

- Original model family: Stable Diffusion 2.1 Base
- Original upstream repository: `stabilityai/stable-diffusion-2-1-base`
- Original developers: Stability AI and collaborators

This mirror republishes the original diffusers-format SD2.1 Base backbone without additional fine-tuning in this repository.

## Model Details

This repository contains the standard diffusers components required to load the SD2.1 Base backbone:

- `model_index.json`
- `feature_extractor/`
- `scheduler/`
- `text_encoder/`
- `tokenizer/`
- `unet/`
- `vae/`

The checkpoint can be loaded directly with Hugging Face Diffusers for text-to-image or image-to-image inference.

## Intended Use

This repository is intended for lawful research, reproducible inference, archival access, and downstream pipelines that comply with the CreativeML Open RAIL++-M license.

Typical uses include:

- text-to-image generation
- image-to-image generation
- backbone loading for training-free downstream pipelines such as STHCD

## Out-Of-Scope Use

Do not use this model for unlawful, harmful, deceptive, privacy-violating, or otherwise restricted purposes. The binding use restrictions are defined by the CreativeML Open RAIL++-M license, especially Attachment A in [LICENSE-MODEL](./LICENSE-MODEL).

## Diffusers Usage

```python
import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "STHCD/stable-diffusion-2-1-base",
    torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")

image = pipe("a satellite photo of a coastal city at sunrise").images[0]
image.save("sd21_example.png")
```

For image-to-image style loading:

```python
import torch
from diffusers import StableDiffusionImg2ImgPipeline

pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
    "STHCD/stable-diffusion-2-1-base",
    torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
```

## Limitations And Bias

This checkpoint inherits the limitations of the original Stable Diffusion 2.1 Base model and the data used to train it. Outputs may reflect social bias, factual errors, unsafe content, or domain mismatch. The model should not be used in safety-critical, deceptive, surveillance, or rights-infringing settings.

## Use In STHCD

This SD2.1 Base backbone is used by the training-free STHCD framework described in:

**A Spatio-Temporal Hierarchical Diffusion Framework for Training-Free Perceptual Remote Sensing Image Compression**  
IEEE Transactions on Geoscience and Remote Sensing, 2026

STHCD uses SD2.1 Base as a perceptual diffusion backbone for remote sensing image compression. The STHCD project is independent and is not affiliated with Stability AI.

## Citation

If you use this mirror or the downstream STHCD workflow, please cite the original Stable Diffusion / Latent Diffusion work together with the STHCD paper when appropriate.

```bibtex
@inproceedings{rombach2022high,
  title={High-Resolution Image Synthesis with Latent Diffusion Models},
  author={Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bjorn},
  booktitle={CVPR},
  year={2022}
}
```

```bibtex
@article{cheng2026spatio,
  title={A Spatio-Temporal Hierarchical Diffusion Framework for Training-Free Perceptual Remote Sensing Image Compression},
  author={Cheng, Yangxuan and Meng, Fanyang and Qi, Hao and Shen, Han and Zhang, Zhongqiang and Wang, Ye and Liang, Yongsheng},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2026},
  volume={64},
  number={},
  pages={1-17},
  publisher={IEEE},
  doi={10.1109/TGRS.2026.3675642}
}
```