Junjie Li · Congyang Ou · Haokui Zhang · Guoting Wei · Shengqin Jiang · Ying Li · Chunhua Shen

Given a PAN–LRMS image pair, SALAD-Pan fine-tunes a pre-trained diffusion model to generate a HRMS.

• [02/01/2026] Code will be released soon !

• Setup
• Usage
• Results
• Citation
• Shoutouts

git clone https://github.com/JJLibra/SALAD-Pan.git
cd SALAD-Pan

conda create -n saladpan python=3.10 -y
conda activate saladpan

# This project depends on a modified local version of `diffusers` under `./diffusers`.
cd diffusers
pip install -e .

cd ..
pip install -r requirements.txt

And initialize an 🤗Accelerate environment with:

accelerate config

Or for a default accelerate configuration without answering questions about your environment

accelerate config default

We provide two-stage checkpoints:

• Stage I (Band-VAE): checkpoints/vae.safetensors (download: Hugging Face)
• Stage II (Latent Diffusion): runs on top of Stable Diffusion in the Band-VAE latent space.
  • Stable Diffusion base: download from Hugging Face (e.g., Stable Diffusion v1-5)
  • Adapters: checkpoints/adapters.pth (download: Hugging Face)

We train the model in two stages.

• Stage I (VAE pretraining)

accelerate launch train_vae.py --config configs/train_vae.yaml

• Stage II (Diffusion + Adapter training)

accelerate launch train_diffusion.py --config configs/train_diffusion.yaml

Note: Tuning usually takes 40k~50k steps, about 1~2 days using eight RTX 4090 GPUs in fp16. Reduce batch_size if your GPU memory is limited.

Once the training is done, run inference:

Coming soon.

Installing xformers is highly recommended for more efficiency and speed on GPUs. To enable xformers, set enable_xformers_memory_efficient_attention=True.

🚨We strongly recommend that you visit this website for a better reading experience.

Table 1. Quantitative results on the WorldView-3 (WV3) dataset. Best results are in bold.

Table 2. Quantitative results on the QuickBird (QB) dataset. Best results are in bold.

Table 3. Quantitative results on the GaoFen-2 (GF2) dataset. Best results are in bold.

Visual comparison on WorldView-3 (WV-3) and QuickBird (QB) dataset at reduced resolution. Visual comparison on WorldView-3 (WV-3) and QuickBird (QB) dataset at full resolution.

Latency is reported as mean ± std over 10 runs (warmup=3), batch size=1, evaluated on the QB dataset under the reduced-resolution (RR) protocol, on an RTX 4090 GPU.

If you make use of our work, please cite our paper.

@article{li2026saladpan,
  title={SALAD-Pan: Sensor-Agnostic Latent Adaptive Diffusion for Pan-Sharpening},
  author={Junjie Li and Congyang Ou and Haokui Zhang and Guoting Wei and Shengqin Jiang and Ying Li and Chunhua Shen},
  journal={arXiv preprint arXiv:2602.04473},
  year={2026}
}

• Built with 🤗 Diffusers. Thanks for open-sourcing !
• The interactive demo is powered by 🤗 Gradio. Thanks for open-sourcing !