RadioFlow 🚀📡

Flow‑Matching for Lightning‑Fast, High‑Fidelity Radio‑Map Generation

---

✨ Why RadioFlow?

RadioFlow is a lightweight, ultra-fast generative model tailored for high-fidelity radio map construction. Compared to existing baselines like diffusion-based and UNet-based methods, it delivers significantly better visual quality, drastically reduced inference time, and an exceptionally compact model size—especially with the edge-friendly RadioFlow-Lite variant. Powered by Conditional Flow Matching, spatial attention UNet, and classifier-free guidance, it achieves state-of-the-art performance with a single-step ODE solver, completely bypassing the costly iterative denoising used in diffusion models.

The framework features a modular design with:

• 🧱 Flexible UNet-based architecture and attention modules
• 🧠 A training pipeline supporting mixed precision, EMA, and real-time visualization
• ⚙️ RadioFlow can be seamlessly scaled down to a lightweight version for edge and embedded devices

From noise to signal map in just one deterministic step. 🚀

📄 Paper: arXiv:2510.09314

🚀 Quick Start

1. Environment Setup

Install dependencies:

pip install -r requirements.txt

2. Dataset

• RadioMapSeer
  Download link

• RadioMap3DSeer
  Download link

Pretrained checkpoints: BaiduNetDisk

3. Training

1. Open config.py and set:
   • data_dir: path to your dataset
   • model_size: 'lite' or 'large' (default: 'lite')
   • training hyperparameters (e.g., learning rate, batch size, number of epochs)
2. Choose the appropriate data loader:
   • RadioUNet_c for the RadioMapSeer dataset
   • RadioMap3Dset for the RadioMap3DSeer dataset
3. Launch training:

   # Train with Lite model (default)
   python train.py
   
   # Train with Large model
   python train.py --model_size large

4. Testing

• SRM evaluation (Lite model):

  python test.py --checkpoint SRM_Lite.pt --task srm

• SRM evaluation (Large model):

  python test.py --checkpoint SRM_Large.pt --task srm --model_size large

• DRM evaluation (Lite model):

  python test.py --checkpoint DRM_Lite.pt --task drm

• DRM evaluation (Large model):

  python test.py --checkpoint DRM_Large.pt --task drm --model_size large

⚠️ Important: Ensure the --model_size parameter matches your checkpoint. Use --model_size large for *_Large.pt checkpoints.

5. Visualization

1. In config.py, configure the VizConfig class to specify visualization options.
2. Run the visualization script:

   # Visualize SRM with Lite model
   python viz.py --task srm
   
   # Visualize SRM with Large model
   python viz.py --task srm --model_size large
   
   # Visualize DRM
   python viz.py --task drm

📝 Reproducing Paper Results

🧪 Task	📦 Dataset	📉 NMSE ↓	🔊 PSNR ↑	📏 RMSE ↓	🧠 SSIM ↑
SRM	RadioMapSeer	0.0023	39.83 dB	0.0103	0.9249
DRM	RadioMapSeer	0.0028	39.37 dB	0.0108	0.9236
SRM	RadioMap3DSeer	0.0496	26.87 dB	0.0458	0.7377

📊 Visual Gallery

DRM Flow (ours) vs RadioUNet	SRM Flow (ours) vs RadioUNet
Fig. 1: DRM Flow comparison	Fig. 2: SRM Flow comparison

DRM Task: CFG Scale Comparison	SRM Task: CFG Scale Comparison
Fig. 3: DRM map outputs under different CFG scale settings	Fig. 4: SRM map outputs under different CFG scale settings

Fig. 5: Quantitative comparison of NMSE, PSNR, RMSE, Time,and Params for RadioFlow against other methods.