RadioDiff-FS: Physics-Informed Manifold Alignment in Few-Shot Diffusion Models for High-Fidelity Radio Map Construction

July 4, 2026 · View on GitHub

📡 Welcome to the RadioDiff Family

Radio map construction via generative diffusion models — UNIC Lab, Xidian University

🔷 Base Backbone

RadioDiff — The foundational diffusion model for radio map construction.   📄 Paper  |  💻 Code  |  IEEE TCCN

🔬 Physics-Informed Extensions

RadioDiff-k² — PINN-enhanced diffusion guided by the Helmholtz equation.   📄 Paper  |  💻 Code  |  IEEE JSAC

iRadioDiff — Indoor radio map construction with physical information integration.   📄 Paper  |  💻 Code  |  IEEE ICC  Best Paper

⚡ Efficiency & Dynamics

RadioDiff-Turbo — Efficiency-enhanced RadioDiff for accelerated inference.   📄 Paper  |  INFOCOM Workshop

RadioDiff-Flux — Adaptive reconstruction under dynamic environments and base station location changes.   📄 Paper  |  IEEE TCCN

🌐 Extended Scenarios

RadioDiff-3D — 3D radio map construction with the UrbanRadio3D dataset.   📄 Paper  |  💻 Code  |  IEEE TNSE

RadioDiff-FS — Few-shot learning for radio map construction with limited measurements.   📄 Paper  |  💻 Code  |  IEEE IoTJ

📶 Sparse Measurement & Localization

RadioDiff-Inverse — Sparse measurement-based radio map recovery for ISAC applications.   📄 Paper  |  💻 Code  |  IEEE TWC

RadioDiff-Loc — Sparse measurement-based NLoS localization using diffusion models.   📄 Paper  |  arXiv

📚 For a comprehensive categorized overview of radio map research, visit Awesome-Radio-Map-Categorized.

Before Starting

  1. install torch
conda create -n radiodiff python=3.9
conda avtivate radiodiff
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
  1. install other packages.
pip install -r requirement.txt
  1. prepare accelerate config.
accelerate config # HOW MANY GPUs YOU WANG TO USE.

Prepare Data

We used the RadioMapSeer dataset for model training and testing.
  • The data structure should look like:
|-- $RadioMapSeer
|   |-- gain
|   |-- |-- carsDPM
|   |-- |-- |-- XXX_XX.PNG
|   |-- |-- |-- XXX_XX.PNG
|   ...
|   |-- png
|   |-- |-- buildings_complete
|   |-- |-- |-- XXX_XX.PNG
|   |-- |-- |-- XXX_XX.PNG
|	...

:tada: Training

  1. train the first stage model (AutoEncoder):
accelerate launch train_vae.py --cfg ./configs/first_radio.yaml
  1. train latent diffusion-edge model:
accelerate launch train_cond_ldm.py --cfg ./configs/radio_train.yaml
  1. fine-tune the pretrained model on IRT4 few-shot data:
accelerate launch train_cond_ldm_finetune.py --cfg ./configs/radio_train_irt4_finetune.yaml

V. Inference.

python sample_cond_ldm.py --cfg ./configs/radio_test_finetune.yaml