Site-Specific MIMO Channel Generation via Diffusion and Flow Matching: Fidelity, Efficiency, and Downstream Utility

Link to paper: https://arxiv.org/abs/2606.20098

Authors: Sina Beyraghi, Masoud Sadeghian, Firdous Bin Ismail, Angel Lozano, Paul Almasan, and Giovanni Geraci

Contact: Sina Beyraghi (mohammadsina.beyraghi@telefonica.com)

Abstract

This paper explores the use of generative models to synthesize high-quality, site-specific multiple-input multiple-output (MIMO) channel data, addressing the high cost of the extensive measurement campaigns required to acquire real-world data for AI-native wireless networks. Two location-conditioned generative paradigms are compared: a conditional denoising diffusion implicit model (cDDIM), and a conditional flow matching model (cFMM). Both these models generate MIMO channel matrices conditioned on user coordinates, to preserve the spatial structure of the deployment site. The approaches are evaluated across three dimensions: statistical fidelity (including beam consistency and effective rank), generation efficiency, and utility in downstream tasks such as channel-state information compression and beam alignment. Results across diverse propagation scenarios (28 GHz and 3.5 GHz, both line-of-sight and non-line-of-sight) demonstrate that both models accurately capture site-specific characteristics, even when trained on scarce ground-truth data. Notably, cFMM achieves a quality comparable to cDDIM with roughly an order of magnitude less inference time. Augmenting scarce site-specific datasets with these synthetic channels yields hefty performance gains in downstream physical layer tasks compared to using scarce data alone or stochastic channels.

Citation

If you use these models or datasets, please cite:

@misc{beyraghi2026sitespecificmimochannelgeneration,
      title={Site-Specific MIMO Channel Generation via Diffusion and Flow Matching: Fidelity, Efficiency, and Downstream Utility}, 
      author={Sina Beyraghi and Masoud Sadeghian and Firdous Bin Ismail and Angel Lozano and Paul Almasan and Giovanni Geraci},
      year={2026},
      eprint={2606.20098},
      archivePrefix={arXiv},
      primaryClass={cs.IT},
      url={https://arxiv.org/abs/2606.20098}, 
}

Repository Structure

GenAI_Channel_Modeling/
├── cDDIM_and_cFMM/          ← Generative model training & inference
├── Channel_Sionna_RT_Github/ ← Dataset generation with Sionna RT ray tracing
├── CRNet_Github/             ← Downstream task: CSI compression
├── DLGF_Github/              ← Downstream task: beam alignment
└── Effective_Rank_Github/    ← Evaluation metric: effective rank analysis

cDDIM_and_cFMM/ — Generative Models

Training and inference scripts for the two generative models proposed in the paper:

• cDDIM — conditional Denoising Diffusion Implicit Model
• cFMM — conditional Flow Matching Model

Both models support LoS-only and mixed LoS+NLoS propagation scenarios. cDDIM must be trained before cFMM to ensure consistent train/val/test dataset splits.

Channel_Sionna_RT_Github/ — Dataset Generation

Scripts to generate large-scale MIMO channel datasets from 3D radio environments using NVIDIA Sionna RT. Produces .npz files containing OFDM channel tensors and optional ray-tracing parameters (AoA, AoD, delays, CIR). These are the datasets used to train the generative models.

CRNet_Github/ — Downstream Task: CSI Compression

Extends CRNet to study how synthetically generated channels (from cDDIM, cFMM, or 3GPP stochastic models) can augment limited real channel measurements for CSI feedback compression training. Sweeps over the number of real training samples to quantify the benefit of synthetic augmentation.

DLGF_Github/ — Downstream Task: Beam Alignment

Extends the DL-GF grid-free beam alignment framework to support ray-tracing, 3GPP stochastic, and generative-model (cDDIM/cFMM) channel datasets. Includes combined pretraining-on-synthetic + finetuning-on-real training strategies for UPA antenna configurations.

Effective_Rank_Github/ — Evaluation: Effective Rank

Tools for computing and visualising the effective rank of UPA MIMO channels in beamspace, and comparing generative model outputs against a ground-truth reference using the Wasserstein-1 distance.

---

How to use the code

Cloning

Open a terminal and execute the following commands to download the github repository.

git clone https://github.com/Telefonica-Scientific-Research/GenAI_Channel_Modeling
cd GenAI_Channel_Modeling 

---

Related resources

• Trained models: GenAI_Channel_Modeling_Models
• Datasets: GenAI_Channel_Modeling_Datasets

---