license: mit
Model Card — Site-Specific MIMO Channel Generation via Diffusion and Flow Matching: Fidelity, Efficiency, and Downstream Utility
Link to paper (to be updated): [TBC]
Authors: Sina Beyraghi, Masoud Sadeghian, Firdous Bin Ismail, Angel Lozano, Paul Almasan, and Giovanni Geraci
Contact: Sina Beyraghi (mohammadsina.beyraghi@telefonica.com), Paul Almasan
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.
Model overview
Two conditional generative model architectures are provided:
| Abbreviation | Full name | Inference mechanism |
|---|---|---|
| cDDIM | Conditional Denoising Diffusion Implicit Model | Reverse diffusion, n_T = 150 steps |
| cFMM | Conditional Flow Matching Model | Euler integration, steps = 10 |
Both share the same Context U-Net backbone (~15.6 M parameters, n_feat = 256) and are conditioned on 3-dimensional UE coordinates (n_classes = 3). Channels are represented in beamspace as two-channel real tensors of shape (2, 4, 32) (real and imaginary parts; 4 Rx × 32 Tx beams for a 2×2 Rx UPA and 4×8 Tx UPA).
Available checkpoints
Checkpoints are organised under logs/ using the naming convention:
{MODEL}_{dataset}_{freq}_{scenario}_{guide_w}_{N_train}_{date}/
where N_train is the number of real training samples used.
cDDIM — 3.5 GHz, LoS
N_train |
Directory |
|---|---|
| 200 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_200_14_05_2026_10_19/ |
| 500 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_500_19_05_2026_09_32/ |
| 1 000 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_1000_19_05_2026_09_33/ |
| 2 000 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_2000_19_05_2026_09_46/ |
| 5 000 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_5000_19_05_2026_10_00/ |
| 10 000 | logs/CDDIM_sina_dataset_3.5GHz_LoS_0.0_10000_20_05_2026_09_55/ |
cDDIM — 3.5 GHz, NLoS + LoS
N_train |
Directory |
|---|---|
| 200 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_200_15_05_2026_14_55/ |
| 500 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_500_19_05_2026_11_51/ |
| 1 000 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_1000_19_05_2026_11_57/ |
| 2 000 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_2000_19_05_2026_11_57/ |
| 5 000 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_5000_19_05_2026_11_58/ |
| 10 000 | logs/CDDIM_sina_dataset_3.5GHz_NLoS+LoS_0.0_10000_19_05_2026_11_58/ |
cDDIM — 28 GHz, LoS
N_train |
Directory |
|---|---|
| 200 | logs/CDDIM_sina_dataset_28GHz_LoS_0.0_200_13_05_2026_15_07/ |
| 500 | logs/CDDIM_sina_dataset_28GHz_LoS_0.0_500_28_05_2026_09_33/ |
| 1 000 | logs/CDDIM_sina_dataset_28GHz_LoS_0.0_1000_27_05_2026_08_52/ |
cFMM — 3.5 GHz, LoS
N_train |
Directory |
|---|---|
| 200 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_200_14_05_2026_10_21/ |
| 500 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_500_19_05_2026_12_22/ |
| 1 000 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_1000_19_05_2026_12_23/ |
| 2 000 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_2000_19_05_2026_12_23/ |
| 5 000 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_5000_19_05_2026_13_10/ |
| 10 000 | logs/FMM_sina_dataset_3.5GHz_LoS_0.0_10000_20_05_2026_09_57/ |
cFMM — 3.5 GHz, NLoS + LoS
N_train |
Directory |
|---|---|
| 200 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_200_15_05_2026_14_57/ |
| 500 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_500_19_05_2026_14_28/ |
| 1 000 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_1000_19_05_2026_14_28/ |
| 2 000 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_2000_19_05_2026_14_29/ |
| 5 000 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_5000_19_05_2026_14_29/ |
| 10 000 | logs/FMM_sina_dataset_3.5GHz_NLoS+LoS_0.0_10000_19_05_2026_14_30/ |
cFMM — 28 GHz, LoS
N_train |
Directory |
|---|---|
| 200 | logs/FMM_sina_dataset_28GHz_LoS_0.0_200_13_05_2026_15_08/ |
| 500 | logs/FMM_sina_dataset_28GHz_LoS_0.0_500_28_05_2026_09_34/ |
| 1 000 | logs/FMM_sina_dataset_28GHz_LoS_0.0_1000_27_05_2026_08_55/ |
Checkpoint contents
Each model directory contains:
| File | Description |
|---|---|
model.pth |
PyTorch state-dict of the trained model |
training_config.txt |
Hyperparameters used during training |
training_log.txt |
Loss curves and validation metrics logged during training |
indices.npy |
Shuffled dataset indices defining the train/val/test split |
train.npy / val.npy / test.npy |
Pre-processed channel arrays for each split |
train_coords.npy / val_coords.npy / test_coords.npy |
Corresponding UE coordinates |
Important: The
indices.npyfile fixes the data split. cFMM checkpoints reuse the indices from the corresponding cDDIM run to ensure identical splits across both models.
Downloading the checkpoints
git clone https://huggingface.co/PaulAlm/GenAI_Channel_Modeling_Models
cd GenAI_Channel_Modeling_Models
unzip logs.zip
Running inference
After downloading, set the save_dir variable in the inference script to the desired model directory and run:
# cDDIM — LoS
python infer_cDDIM_LoS.py generate
# cFMM — LoS
python infer_cFMM_LoS.py generate
Full instructions are in the code repository.
Training details
| Hyperparameter | cDDIM | cFMM |
|---|---|---|
| Epochs | 3 000 | 2 000 |
| Batch size | 100 | 100 |
| Learning rate | 1 × 10⁻⁴ | 1 × 10⁻⁴ |
| Inference steps | 150 (DDIM) | 10 (Euler) |
| Conditioning | 3D UE coordinates | 3D UE coordinates |
| Guidance weight | 0.0 | 0.0 |
| Model parameters | ~15.6 M | ~15.6 M |
Datasets
The corresponding channel datasets are available at:
https://huggingface.co/datasets/PaulAlm/GenAI_Channel_Modeling_Datasets
Related resources
- Code repository: GenAI_Channel_Modeling
- Datasets: GenAI_Channel_Modeling_Datasets
Citation
If you use these models, please cite:
TBC.