Spectrum-aware Highly-dynamic Adaptation for Resolution Promotion in Remote Sensing Synthesis

Training-free large-scale remote sensing text-to-image synthesis with an RS-adapted FLUX prior and dynamic positional adaptation.

🌍 Overview

SHARP is a training-free resolution promotion framework for remote sensing text-to-image synthesis. Built on top of an RS-adapted FLUX prior, it applies stronger positional extrapolation during early layout formation and progressively relaxes it during late detail recovery, enabling robust large-scale generation while preserving the dense high-frequency structures that are especially important in remote sensing imagery.

In short: SHARP helps generate sharper, larger, and more structure-faithful remote sensing images without extra training.

The fine-tuned checkpoint is available at: BxuanZ/FLUX-RS

🔥 Highlights

• ✨ Training-free resolution promotion for remote sensing image synthesis
• 🛰️ Spectrum-aware dynamic positional adaptation aligned with diffusion denoising
• 📐 Resolution-agnostic generation across square and rectangular high resolutions
• ⚡ FLUX-based implementation powered by an RS-specialized generative prior
• 🧩 Single-GPU and multi-GPU entry points for both quick demos and batch evaluation

🖼️ Method Snapshot

The figure below gives a quick look at the SHARP framework and its design intuition:

📢 Current Status

• ✅ SHARP inference code
• ✅ Fine-tuned RS-FLUX weights
• ❌ Training data

🧭 Repository Map

Click to expand the project layout

SHARP/
├── assets/
│   └── logo.png
├── checkpoints/
│   └── .gitkeep
├── docs/
│   └── structure.png
├── flux/
│   ├── pipeline_flux.py
│   └── transformer_flux.py
├── LICENSE
├── README.md
├── requirements.txt
├── rs_t2i_eval_prompts_100.txt
├── run_sharp.py
├── run_sharp.sh
└── run_sharp_multi_gpu.py

📦 What Is Included

File	Description
run_sharp.py	Official SHARP single-GPU generation entry
run_sharp_multi_gpu.py	Multi-GPU batch launcher for multi-scale generation
run_sharp.sh	Thin shell wrapper around run_sharp.py
flux/pipeline_flux.py	SHARP pipeline implementation
flux/transformer_flux.py	SHARP transformer with dynamic positional adaptation
rs_t2i_eval_prompts_100.txt	Example prompt list for quick evaluation
checkpoints/	Default place for storing local checkpoints

🛠 Installation

1. Create the environment

conda create -n sharp python=3.10
conda activate sharp

2. Install dependencies

pip install -r requirements.txt

3. Main dependencies

torch
torchvision
diffusers
transformers
accelerate
sentencepiece

📥 Checkpoints

The recommended checkpoint is the RS-adapted FLUX model hosted on Hugging Face:

• FLUX-RS: https://huggingface.co/BxuanZ/FLUX-RS

By default, SHARP looks for checkpoints under:

checkpoints/

You can use either workflow below:

• Place the downloaded FLUX-RS model directory under checkpoints/, then run SHARP without changing --ckpt_path
• Pass the checkpoint directory explicitly with --ckpt_path /path/to/your_checkpoint_dir

A valid checkpoint directory should contain a transformer/ subfolder, for example:

checkpoints/<your_checkpoint_dir>/

🚀 Quick Start

1. Generate from a single prompt

bash run_sharp.sh \
  --prompt "A satellite image of a rural market town with dense shop blocks, a bus station, surrounding crop fields, narrow feeder roads, and mixed residential and commercial parcels." \
  --width 1024 \
  --height 1024

2. Generate from a prompt file

python run_sharp.py \
  --prompt_file rs_t2i_eval_prompts_100.txt \
  --width 1024 \
  --height 1536 \
  --ckpt_path checkpoints/<your_checkpoint_dir> \
  --save_prefix sharp_eval \
  --out_dir sharp_outputs

3. Launch multi-GPU evaluation

python run_sharp_multi_gpu.py \
  --gpus 0 1 2 \
  --prompt_file rs_t2i_eval_prompts_100.txt \
  --ckpt_path checkpoints/<your_checkpoint_dir> \
  --scales 1024x1024 1764x1764 1024x1536 1920x1024

⚙️ Helpful Notes

• SHARP uses one fixed inference path, so no method switch is needed.
• The default single-GPU output directory is sharp_outputs/.
• The default multi-GPU output directory is sharp_outputs_eval/.
• run_sharp_multi_gpu.py will auto-detect GPUs when --gpus is not provided.
• If --ckpt_path points to checkpoints/, SHARP can auto-discover the model only when exactly one valid checkpoint directory exists there.
• For FLUX latent packing, image sizes are ideally divisible by 16. Otherwise, the effective generated size may be rounded down internally.
• --skip_existing is useful for resuming long batch jobs without re-generating finished outputs.

🧪 Script Entry Points

bash run_sharp.sh
python run_sharp.py --help
python run_sharp_multi_gpu.py --help

These entry points were sanity-checked with --help.

🙏 Acknowledgement

SHARP is built on an RS-adapted FLUX prior and includes custom remote sensing generation utilities for practical large-resolution synthesis.

📚 Citation

If you find SHARP useful in your research, please cite our paper:

@misc{zhao2026sharpspectrumawarehighlydynamicadaptation,
      title={SHARP: Spectrum-aware Highly-dynamic Adaptation for Resolution Promotion in Remote Sensing Synthesis},
      author={Bingxuan Zhao and Qing Zhou and Chuang Yang and Qi Wang},
      year={2026},
      eprint={2603.21783},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.21783},
}

📌 License

This project is released under the MIT License.