• Continuous-Scale SR: Support arbitrary scale factors from 1× to 8× with a single model
• Two-Stage Latent Diffusion: Efficient super-resolution in compressed latent space
• Few-Step Inference: Only 4 diffusion steps required for high-quality results
• Remote Sensing Optimized: Designed specifically for aerial and satellite imagery

1. Clone repo

git clone https://github.com/MoooJianG/LDCSR.git

2. Install dependencies

conda create -n LDCSR python=3.10
conda activate LDCSR
pip install -r requirement.txt

We support AID, DOTA, and DIOR out‑of‑the‑box. Download HR images using the official links below (or your own data) and generate LR/HR pairs via bicubic down‑sampling.

Supported Datasets:

• AID - Aerial Image Dataset
• DOTA - Dataset for Object Detection in Aerial Images
• DIOR - Detection In Optical Remote sensing images

# Example: split AID
python data/prepare_split.py --split_file AID_split.pkl --data_path dataset/RawAID --output_path dataset/AID

Custom datasets should replicate the following folder structure:

└── dataset
    └── YourData
        ├── Train
        |   ├── HR
        |   └── LR
        ├── Test
        └── Val

# First-stage
python train.py --config configs/first_stage_kl_v6.yaml
# Second-stage
python train.py --config configs/second_stage_van_v4.yaml

python test.py --checkpoint path/to/checkpoint.ckpt --datasets AID --scales 4

Download pretrained models and dataset splits from Google Drive.

Checkpoints:

• first_stage/ - First stage VAE model
• second_stage/ - Second stage diffusion model

Dataset Splits:

• AID_split.pkl - AID dataset split
• DIOR_400test.pkl - DIOR test split
• DOTA_400test.pkl - DOTA test split
• NWPU_100test.pkl - NWPU test split

LDCSR adopts a two-stage latent diffusion framework:

Stage 1 - Variational Autoencoder (VAE)

• Encodes HR images into a compact 4-channel latent space (8× spatial compression)
• GAPEncoder/GAPDecoder with scale-aware decoding for continuous-scale reconstruction
• Trained with reconstruction loss, KL divergence, and adversarial loss (LPIPS + Discriminator)

Stage 2 - Latent Diffusion Model

• Performs super-resolution directly in the latent space
• UNet denoiser conditioned on: LR features, scale factor, and output size embeddings
• Only 4 diffusion steps for efficient inference

Quantitative comparison on AID dataset for integer scale factors (×2, ×4, ×8). Best results in bold.

Fixed-Scale Methods:

Continuous-Scale Methods:

Our method achieves the best perceptual quality (LPIPS, FID) among continuous-scale methods while maintaining competitive PSNR. See the paper for full results on AID, DOTA, and DIOR datasets.

Key parameters in config files:

First Stage (configs/first_stage_kl_v6.yaml)

• embed_dim: Latent space dimension (default: 4)
• ch_mult: Channel multipliers for encoder/decoder
• disc_start: Step to enable discriminator training

Second Stage (configs/second_stage_van_v4.yaml)

• timesteps: Number of diffusion steps (default: 4)
• linear_start/end: Noise schedule parameters
• scale_by_std: Whether to normalize latent by std

If you have any questions or suggestions, feel free to contact me.

Email：20220119004@bfsu.edu.cn

@ARTICLE{11006698,
  author={Wu, Hanlin and Mo, Jiangwei and Sun, Xiaohui and Ma, Jie},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Latent Diffusion, Implicit Amplification: Efficient Continuous-Scale Super-Resolution for Remote Sensing Images}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Diffusion models;Training;Image synthesis;Noise reduction;Visualization;Decoding;Computational modeling;Remote sensing;Image reconstruction;Autoencoders;Remote sensing;super-resolution;latent diffusion;continuous-scale},
  doi={10.1109/TGRS.2025.3571290}}

This project is released under the MIT License.

This work is built upon several excellent open-source projects:

• Latent Diffusion Models
• PyTorch Lightning
• LPIPS