RadioDiff-3D: Volumetric Diffusion for Radio Maps

• The paper introduces RadioDiff-3D as a latent diffusion model that generates volumetric radio maps, capturing metrics including pathloss, DoA, and ToA.
• The framework extends traditional 2D pathloss prediction to a 3D environment by leveraging a 3D convolutional U-Net and conditional guidance from environment metadata and sparse observations.
• RadioDiff-3D establishes a foundational benchmark for environment-aware, multi-modal radio mapping, fostering advancements in 6G, aerial networking, and volumetric channel prediction.

RadioDiff-3D is a conditional diffusion framework for 3D radio map generation introduced together with UrbanRadio3D, a large-scale, high-resolution dataset for environment-aware wireless communication. It extends earlier radio-map learning from fixed-height 2D pathloss prediction to volumetric generation over multiple receiver heights and multiple channel descriptors, notably pathloss, direction of arrival (DoA), and time of arrival (ToA). The framework is formulated for both radiation-aware settings, where transmitter metadata are known, and radiation-unaware settings, where sparse spatial observations guide reconstruction (Wang et al., 16 Jul 2025).

1. Scope, motivation, and conceptual position

RadioDiff-3D was proposed against a background in which radio-map construction was dominated by two imperfect families of methods. Physics-based solvers, including full-wave EM methods and ray tracing, offer high fidelity but incur substantial latency. Learning-based methods are faster, but most prior systems predicted only 2D pathloss maps at a single receiver height, typically omitting vertical variation and higher-order channel descriptors such as DoA and ToA. The stated motivation is that 6G environment-aware communication, XL-MIMO operation, aerial networking, volumetric coverage planning, and angle- or delay-aware control require richer spatial channel knowledge than single-plane pathloss images can provide (Wang et al., 16 Jul 2025).

Within the diffusion lineage, RadioDiff-3D can be read as the volumetric generalization of the earlier RadioDiff paradigm. Related 2D work such as RadioDiff-$k^2$ remained centered on $N \times N$ pathloss images and introduced Helmholtz-informed singularity guidance rather than volumetric generation, which underscores the methodological shift represented by RadioDiff-3D from planar RM synthesis to 3D multi-metric generation (Wang et al., 22 Apr 2025).

A common misconception is that RadioDiff-3D is only a 3D dataset paper or only a 3D U-Net benchmark. In fact, the contribution is dual: UrbanRadio3D supplies the data regime that earlier methods lacked, and RadioDiff-3D supplies a conditional diffusion benchmark specifically intended for 3D RM construction (Wang et al., 16 Jul 2025).

2. UrbanRadio3D dataset and the meaning of “3D×3D”

UrbanRadio3D is a large-scale, high-resolution 3D radio-map dataset generated with WinProp using the Dominant Path Model. Each urban region covers $256 \times 256\ \text{m}^2$ at $1\ \text{m}$ horizontal resolution and includes 20 receiver-height layers from 1 m to 20 m in 1 m increments. The dataset contains 701 distinct urban regions derived from cities including Kara, Berlin, Glasgow, Ljubljana, London, and Tel Aviv. Building heights range from $6.6\ \text{m}$ to $19.8\ \text{m}$, and 200 transmitter locations are sampled per urban map. The resulting simulation count is $701 \times 200 \times 20 = 2.84$ million; the paper also reports a total dataset scale of 11.2M when counting channel modalities and images (Wang et al., 16 Jul 2025).

The radio modalities are pathloss, DoA azimuth, DoA elevation, and ToA, with auxiliary representations including building segmentation, grayscale building-height maps, transmitter location maps, and propagation-ray data. The dataset comparison in the paper emphasizes that UrbanRadio3D provides 20 height layers, $1\ \text{m}$ height resolution, real building shape, real building height, and directional/time-of-arrival labels, while also being over $37\times$ larger than prior datasets (Wang et al., 16 Jul 2025).

The paper’s phrase “3D×3D dataset” refers to two simultaneous extensions. First, the environment is genuinely 3D in geometry, with realistic urban height variation. Second, the radio representation spans 3D space and multiple metric families—pathloss, DoA, and ToA—rather than a single scalar 2D field. In operational terms, the learning target is a tensor

$$\mathcal{R} \in \mathbb{R}^{H \times W \times D \times C},$$

where $N \times N$0 are spatial dimensions and $N \times N$1 indexes channel modalities (Wang et al., 16 Jul 2025).

3. Formal problem statement and diffusion formulation

RadioDiff-3D defines two settings. In the radiation-aware setting, the inputs are a 3D environment occupancy grid

$N \times N$2

and transmitter metadata

$N \times N$3

with target distribution

$N \times N$4

In the radiation-unaware setting, the transmitter is unknown or non-cooperative, and the condition is instead a sparse observation set

$N \times N$5

yielding

$N \times N$6

This bifurcation is central: the same framework covers environment-to-map generation and sparse-observation-conditioned reconstruction (Wang et al., 16 Jul 2025).

The paper presents RadioDiff-3D as a latent diffusion framework. A variational encoder maps the radio map to latent space,

$N \times N$7

after which a standard forward diffusion process is applied: $N \times N$8 Equivalently,

$N \times N$9

The denoiser predicts noise conditioned on side information $256 \times 256\ \text{m}^2$0, with training objective

$256 \times 256\ \text{m}^2$1

The paper also restates the standard DDPM preliminaries in $256 \times 256\ \text{m}^2$2 notation and includes a DDIM-style deterministic sampling formulation for accelerated inference (Wang et al., 16 Jul 2025).

For radiation-unaware inference, sparse observations are injected through a guidance term in latent space: $256 \times 256\ \text{m}^2$3 where $256 \times 256\ \text{m}^2$4 is a sparse RM interpolated to output resolution. This makes sparse observation consistency an explicit denoising-time correction rather than a purely input-level cue (Wang et al., 16 Jul 2025).

4. Architecture, representation, and conditioning

The denoising backbone is a 3D convolutional U-Net with residual blocks, attention layers, and encoder–decoder skip connections. The architectural rationale is straightforward: 3D convolutions model dependencies jointly across width, height, and altitude, and therefore can capture inter-slice continuity, height-varying LoS/NLoS transitions, vertical angular variation, and cross-height correlations induced by building geometry. The denoiser processes a noisy 3D latent tensor and outputs either the predicted noise or a denoised latent estimate (Wang et al., 16 Jul 2025).

Conditioning is denoted $256 \times 256\ \text{m}^2$5. In the known-transmitter setting,

$256 \times 256\ \text{m}^2$6

The paper states that this conditioning is injected into each residual block of the 3D U-Net through cross-attention or FiLM-like modulation, with attention written as

$256 \times 256\ \text{m}^2$7

Thus the diffusion model is conditional at every denoising step rather than merely initialized by environment metadata (Wang et al., 16 Jul 2025).

The experimental benchmark in the paper uses a reduced 4-slice setup. Without sparse sampling, the practical input tensor is described as

$256 \times 256\ \text{m}^2$8

with three channels corresponding to building segmentation, building height, and transmitter location. With sparse conditioning, the input becomes

$256 \times 256\ \text{m}^2$9

where the fourth channel is a sparse sampling map. In these experiments, the target output is pathloss over the same 4-slice depth (Wang et al., 16 Jul 2025).

The paper also mentions an autoregressive height-wise generation strategy,

$1\ \text{m}$0

introduced as a mechanism for reducing memory cost on larger 3D volumes while preserving vertical continuity. This suggests a hybrid between fully volumetric generation and ordered slice prediction, although the paper does not provide a full ablation of that choice (Wang et al., 16 Jul 2025).

5. Benchmark results, ablations, and what was actually validated

The empirical section is more limited than the conceptual scope of the framework. For RadioDiff-3D itself, the main reported diffusion results are pathloss experiments on the 4-slice subset corresponding to receiver heights 1 m to 4 m. Under the preliminary no-sampling setting with 1000 diffusion steps, the paper reports RMSE 0.0653, NMSE 0.0534, SSIM 0.8309, and PSNR 24.00. Under DDIM sampling with 200 steps and no sparse conditioning, the reported numbers are RMSE 0.1325, NMSE 0.3472, SSIM 0.6453, and PSNR 19.57. With 10% sparse sampling, those improve to RMSE 0.0481, NMSE 0.0550, SSIM 0.8187, and PSNR 29.23, showing that sparse guidance materially sharpens reconstruction in the radiation-unaware setting (Wang et al., 16 Jul 2025).

The DDIM inference-time study reports 2.43 s at 20 steps, 3.6574 s at 30 steps, 6.1171 s at 50 steps, 12.5732 s at 100 steps, and 24.3526 s at 200 steps, exhibiting the expected approximately linear latency–quality tradeoff (Wang et al., 16 Jul 2025).

The most detailed metric tables in the paper are actually for the deterministic 3D-UNet benchmark on ToA, DoA_Azi, and DoA_Ele rather than for RadioDiff-3D itself. For example, under uniform 10% sampling, the 3D-UNet reports ToA RMSE 0.0140, NMSE 0.0010, SSIM 0.9849, and PSNR 37.39; DoA_Azi RMSE 0.0299, NMSE 0.0047, SSIM 0.9677, and PSNR 31.18; and DoA_Ele RMSE 0.0227, NMSE 0.0021, SSIM 0.9783, and PSNR 34.29. These results validate the utility of the dataset and the 3D convolutional setting, but they should not be conflated with a fully comprehensive RadioDiff-3D evaluation across all modalities (Wang et al., 16 Jul 2025).

A recurrent misunderstanding is that RadioDiff-3D was exhaustively validated on all 20 heights and all four output modalities. The paper does not do that. It explicitly narrows the main learning experiments to the 1–4 m subset, and the most complete quantitative tables for ToA and DoA are reported for the 3D-UNet baseline rather than for the diffusion model (Wang et al., 16 Jul 2025).

6. Limitations, later comparative use, and role in the literature

Several limitations are explicit in the paper’s own presentation. First, although UrbanRadio3D contains 20 receiver heights, the reported benchmarked learning tasks use only four slices. Second, full multi-modal joint generation with RadioDiff-3D is not comprehensively quantified. Third, the dataset uses simplified material parameters: all building facades are assigned default WinProp material properties with surface thickness $1\ \text{m}$1, transmission loss $1\ \text{m}$2, reflection loss $1\ \text{m}$3, $1\ \text{m}$4, $1\ \text{m}$5, and $1\ \text{m}$6. Fourth, the environments are effectively static. Fifth, some reverse DDPM and DDIM expressions in the manuscript are typeset incorrectly, and the “latent diffusion” description is not matched by a fully specified encoder–decoder implementation (Wang et al., 16 Jul 2025).

Despite those constraints, RadioDiff-3D quickly became a reference point in subsequent work. PILOT, an autoregressive alternative for unified 2D and 3D radio-map construction, compares directly against RadioDiff-3D on the UrbanRadio3D 1–4 m benchmark and reports that RadioDiff-3D achieves NMSE 0.0534, RMSE 5.03, SSIM 0.8309, PSNR 24.00, and 121.7630 s inference at $1\ \text{m}$7, whereas PILOT volumetric reports NMSE 0.0120, RMSE 2.60, SSIM 0.9290, PSNR 29.92, and 0.0486 s. That comparison established RadioDiff-3D as the canonical volumetric diffusion baseline for later sampling-free generative alternatives (Huang et al., 26 Apr 2026).

In aerial channel-map prediction, a later geometry-aware cross-height CKM study identifies 3D-RadioDiff as “the strongest baseline” and characterizes it as “An Altitude-Conditioned Diffusion Model for 3D Radio Map Construction.” Under that benchmark, 3D-RadioDiff records RMSE 6.937 dB in unseen-scene zero-shot evaluation and 1.221 dB under the easier legacy patch-random protocol. The same paper also discloses one concrete implementation of the baseline: 50 forward diffusion steps, 25 sampling steps at test time, 12 sampling steps during validation, diffusion base channels 32, a beta schedule from $1\ \text{m}$8 to $1\ \text{m}$9, and a diffusion loss combining noise prediction, reconstruction, and gradient terms with weights 1.0, 0.8, and 0.10 (Zeng et al., 1 Jul 2026).

A broader literature trend places RadioDiff-3D between two adjacent developments. One direction keeps diffusion but targets deployment constraints, as in InvDiff-CGM, which is a 2D CGM method with 3D-aware priors and explicitly cites RadioDiff-3D as prior volumetric work (Gao et al., 13 Apr 2026). The other direction abandons iterative denoising in favor of alternative generative orders or explicit physics prompts, as in PILOT’s wavefront autoregression (Huang et al., 26 Apr 2026). This suggests that RadioDiff-3D’s enduring importance is less that it solved 3D radio mapping definitively, and more that it defined the first broadly adopted volumetric diffusion benchmark for the field.

In that sense, RadioDiff-3D occupies a foundational but transitional position. It provided the dataset regime, tensor formalism, and conditional diffusion baseline needed to move radio-map learning from fixed-height 2D pathloss estimation toward volumetric, multi-metric, environment-aware modeling for 6G systems (Wang et al., 16 Jul 2025).