Fine-tuned Stable Diffusion 1.5 backbone for generating synthetic high-resolution SAR amplitude images from text prompts.
Published alongside the upcoming paper "Diffusion-Based SAR Training Data Synthesis Controlled by Spatial Annotations" (Hochstuhl et al., 2026; accepted for GCPR conference 2026).

Database

Fine-tuning was conducted on a SAR–text dataset compiled from high-resolution Capella (X-band) imagery originating from the SpaceNet6 challenge dataset, which covers the Rotterdam harbor area. The dataset contains geocoded, fully polarimetric Ground Range Detected (GRD) amplitude images in dB-scale with a ground sample distance of 0.5 m, along with co-registered optical WorldView-2 imagery (RGB, 0.5 m GSD). To create SAR–text pairs, the imagery was tiled into 512 × 512 patches (~256 × 256 m), and text captions were generated from the corresponding optical patches using a BLIP model fine-tuned on remote sensing imagery (BLIP_RSCID).

Architecture

Component	Details
Base model	Stable Diffusion 1.5
VAE decoder	Adapted to single-channel output
Text encoder	Fine-tuned with a LoRA adapter (`adapter_text_encoder/`)
Output	Single-channel float32 SAR amplitude image

Repository Structure

text

HR-SAR-StableDiffusion/
├── model_index.json
├── unet/
├── vae/                        # Modified: single-channel conv_out
├── text_encoder/
├── tokenizer/
├── scheduler/
├── feature_extractor/
└── adapter_text_encoder/       # LoRA adapter for text encoder
    ├── adapter_config.json
    └── adapter_model.safetensors

Usage

This model can be used directly with the 🤗 diffusers pipeline. The LoRA adapter for the text encoder is included in this repository and must be loaded manually after initialising the pipeline.

Requirements:

torch (CUDA recommended)
diffusers
peft

Text-to-Image Generation

python

import torch
from diffusers import StableDiffusionPipeline
from peft import PeftModel

MODEL_ID = "sylviaHoch/HR-SAR-StableDiffusion"

# Load pipeline
pipeline = StableDiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16
).to("cuda")

# Load LoRA text-encoder adapter
pipeline.text_encoder = PeftModel.from_pretrained(
    pipeline.text_encoder,
    MODEL_ID,
    subfolder="adapter_text_encoder"
)

# Generate
image = pipeline(
    "An aerial view of oil tanks.",
    num_inference_steps=50,
    guidance_scale=3.0,
    output_type="pt"
).images

image is a torch.Tensor of shape (N, C, H, W) with values in the range [0, 1] and dtype float32, where N is the batch size (here 1), C the number of channels (here 1, grayscale), and H and W the native model resolution (512 x 512).

Intended Use

Synthetic SAR data generation.
Research on generative models for remote sensing.

Limitations

Tuned specifically to Capella sensor characteristics based on multilooked, geocoded images in VH polarization; transferability to other sensors and imaging properties is very limited.
Generated images are synthetic and may not fully capture all real SAR image properties.

Citation

If you use this model, please cite:

bibtex

@inproceedings{sar_diffusion_gcpr2026,
  title     = {Diffusion-Based SAR Training Data Synthesis Controlled by Spatial Annotations},
  booktitle = {German Conference on Pattern Recognition (GCPR)},
  year      = {2026},
  note      = {accepted, to be published},
  authors   = {}
}

License

This model is released under CC BY-NC-SA 4.0.
Commercial use is not permitted. Derivatives must be shared under the same license.
See LICENSE for details.