mmWave-Diffusion in Wireless Systems

• mmWave-Diffusion is an umbrella term in mmWave research that combines diffusion-model-based inference with physical propagation phenomena such as diffuse scattering and multipath fading.
• It encompasses methods for channel estimation, beam management, radar enhancement, and vehicular networking, all geared toward improving sparse measurement recovery under non-ideal conditions.
• Current approaches focus on efficient inference techniques and context-conditioned generation to reconcile theoretical models with practical, data-driven constraints in mmWave systems.

Searching arXiv for papers on “mmWave-Diffusion” and closely related diffusion-based mmWave methods. mmWave-Diffusion is an umbrella expression used in several distinct strands of millimeter-wave research. In one strand, it denotes diffusion-model-based inference for channel estimation, beam management, radar enhancement, anomaly localization, RF heatmap synthesis, and physiological sensing. In another, it refers to physical diffusion phenomena—diffuse scattering, diffuse multipath, and two-wave-with-diffuse-power fading—that become prominent because short wavelengths, large apertures, and rough surfaces make mmWave propagation highly sensitive to blockage, surface texture, and off-specular energy. A further usage appears in vehicular networking, where “diffusion” denotes the early adoption phase of mmWave-equipped vehicles rather than a generative model (Hu et al., 21 Oct 2025, Wang et al., 21 Mar 2026, Yaman et al., 2019, Taya et al., 2018).

1. Terminology and scope

The literature does not use mmWave-Diffusion to denote a single standardized framework. Instead, the term spans learned generative methods, physical-wave scattering models, and, in V2X, deployment-stage analysis. This overloading is technically important because the mathematical object called “diffusion” changes across papers: stochastic or deterministic reverse-time generative processes in some works, and angular-delay spreading induced by rough surfaces or random media in others.

Usage of “diffusion”	Technical object	Representative examples
Generative diffusion	Score-based, latent, or discrete denoising model	DiffPace (Hu et al., 21 Oct 2025), respiration sensing (Wang et al., 21 Mar 2026), beam candidate generation (Azarbahram et al., 9 Apr 2026)
Physical diffusion	Diffuse scattering, diffuse multipath, diffuse power	RT-ICM (Yaman et al., 2019), FDTD rough-slab modeling (Bakirtzis et al., 2020), TWDP fading (Zöchmann et al., 2018)
Adoption diffusion	Early penetration of mmWave-equipped vehicles	Distributed A3C relay formation (Taya et al., 2018)

A useful organizing distinction is between distribution learning and propagation diffusion. The former learns a conditional prior over channels, radar spectra, point clouds, or RF heatmaps. The latter models how energy spreads in angle, delay, frequency, or space because of roughness, clutter, and multipath. This suggests that mmWave-Diffusion is best understood as a family of methods for recovering latent structure from measurements whose statistics are strongly shaped by sparse but non-ideal propagation.

2. Diffusion-model priors for communication-side inference

In mmWave and THz UM-MIMO channel estimation, the main inverse problem is severely underdetermined because hybrid beamforming compresses observations through a limited number of RF chains. DiffPace formulates the beamspace problem as

$y=\mathrm{vec}(Y)=\Phi h+n,$

learns a score-based diffusion prior over HPSM channels, and injects that prior into a plug-and-play solver for MAP-like reconstruction (Hu et al., 21 Oct 2025). Its central premise is that predefined sparsifying dictionaries cannot faithfully capture hybrid near-field and far-field structure, whereas a diffusion model trained on the Hybrid Planar–Spherical Model can. The prior step is implemented by a probability-flow ODE update,

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$

and the data-consistency step by a closed-form projection,

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$

The paper reports competitive channel-estimation accuracy, including approximately $-15$ dB NMSE at $10$ dB SNR in mmWave settings, while using about 100 ODE steps rather than 1000+ stochastic diffusion steps, i.e., about 90\% fewer inference steps (Hu et al., 21 Oct 2025).

A related communication-side use of diffusion appears in limited-probing beam management. “Discrete Diffusion for Codebook-Based Beam Candidate Generation” models probing as a sequential, partially observable decision problem in which the serving beam must be drawn from a small probed set (Azarbahram et al., 9 Apr 2026). The method learns a history-conditioned discrete denoising diffusion probabilistic model over beam indices, with a hierarchical Transformer encoder producing the history context and a reverse categorical denoiser generating candidate beams. The forward corruption is categorical,

$$q(x_\tau|x_0)=\bar{\alpha}_\tau\,\mathbf{1}\{x_\tau=x_0\}+\frac{1-\bar{\alpha}_\tau}{K},$$

and the training objective uses sparse soft labels over top-$M$ oracle beams rather than a full ELBO. Online inference oversamples reverse trajectories, then ranks beams by frequency and confidence before selecting the probing set. Numerical analysis shows lower beam-miss probability and lower conditional probe regret under tight probing budgets, with gains most pronounced in low-probing regimes (Azarbahram et al., 9 Apr 2026).

These communication-oriented methods share a common structure. Both replace rigid parametric assumptions with learned conditional distributions, and both use the learned model to populate a feasible set that is subsequently constrained by physics or protocol: measurement consistency in DiffPace, and probe-then-serve constraints in codebook beam management. A plausible implication is that diffusion becomes particularly attractive when mmWave inference is bottlenecked not by raw model capacity, but by combinatorial sparsity, basis mismatch, or aggressive compression.

3. Radar, RF fields, and context-conditioned generation

A major branch of mmWave-Diffusion uses diffusion to enhance, synthesize, or compare radar and RF spatial representations. “Radar-diffusion” defines a forward process by an Ornstein–Uhlenbeck mean-reverting SDE,

$$\mathrm{d}x=\theta_t(\mu-x)\mathrm{d}t+\sigma_t\mathrm{d}w,$$

where $x$ is a LiDAR BEV image and $\mu$ is the radar BEV condition, so that the forward degradation converges toward a radar-like biased terminal distribution rather than zero-mean white noise (Luan et al., 2024). The reverse model is trained with a LiDAR-supervised objective that separately weights target and blank regions,

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$0

to suppress ghost points without simply blanking the scene. On View-of-Delft, the method improves FID_BEV from 247.2 to 118.4 and reduces CD from 0.34 to 0.19; it also improves downstream point-based registration, increasing recall from 88.51\% to 93.10\% (Luan et al., 2024).

A different radar enhancement strategy is “Diffusion-Based mmWave Radar Point Cloud Enhancement Driven by Range Images,” which argues that range images align better with natural image statistics than range-azimuth heatmaps or BEV, allowing transfer from pre-trained image diffusion models (Wu et al., 4 Mar 2025). The method constructs multi-channel radar range images, conditions a diffusion UNet in that image-native domain, and uses deterministic PF-ODE sampling with a Heun solver. On ColoRadar, it reports, for example, Hallways CD 2.247, MHD 0.275, and F-Score 44.0\%, compared with OS-CFAR CD 8.326 and F-Score 20.7\%; on Edgar Mine it reports CD 0.413, MHD 0.041, and F-Score 78.7\% (Wu et al., 4 Mar 2025). The underlying claim is not that diffusion invents arbitrary geometry, but that conditioning in a representation close to human observation improves denoising and densification.

Context-conditioned generation is pushed further in mmAnomaly. That framework combines RGBD context with mmWave Range–Azimuth maps, extracts semantic cues such as clothing material or wall type with lightweight ResNet-18 classifiers, and uses a conditional latent diffusion model to synthesize the expected anomaly-free spectrum for the observed scene (Toha et al., 1 Apr 2026). The reverse model follows standard latent diffusion,

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$1

while anomaly localization is performed by a dual-branch ViT that compares real and generated spectra. Across concealed weapon localization, through-wall intruder localization, and through-wall fall localization, the paper reports up to 94\% F1 score and sub-meter localization error (Toha et al., 1 Apr 2026). Here the diffusion prior functions as a context-conditioned null hypothesis: residuals between measured and generated spectra are treated as anomaly evidence.

At a larger spatial scale, Diffusion$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$2 generates RF heatmaps from 3D point clouds, sparse measurements, and RF metadata by conditioning a latent diffusion backbone on an RF-3D Encoder built from MinkUNet18A, Swin-Transformer-based 2D features, and Fourier embeddings of transmitter location, mesh structure, and frequency (Park et al., 2 Oct 2025). The conditional reverse process is

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$3

with deterministic sampling during inference. The reported median error is 1.20 dB for synthetic mmWave and 1.9 dB for Wi-Fi, with inference around 0.59 s per heatmap and 27x speedup over AutoMS (Park et al., 2 Oct 2025). This suggests that diffusion is also being used as a surrogate for expensive propagation simulation when geometry is available but exhaustive measurement is not.

4. Observation-anchored diffusion for respiration sensing

The paper explicitly titled “mmWave-Diffusion: A Novel Framework for Respiration Sensing Using Observation-Anchored Conditional Diffusion Model” applies diffusion to contactless respiratory waveform reconstruction from 60 GHz radar phase observations (Wang et al., 21 Mar 2026). Its core design choice is to model the residual

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$4

where $$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$5 is the normalized radar phase segment and $$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$6 is the synchronized respiration ground truth, and then inject this residual in the forward process:

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$7

Unlike standard DDPM initialization from Gaussian noise, the reverse chain starts in an observation-consistent neighborhood,

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$8

so that sampling remains close to the measurement manifold.

The denoiser is the Radar Diffusion Transformer, conditioned explicitly on the phase observation and structured to preserve temporal correspondence. Two parallel streams process the reverse variable and the observation; identical non-overlapping patching is combined with dual learnable positional encodings; and a banded-mask multi-head cross-attention constrains interactions to temporally neighboring patches,

$$z_i=h_{t_i}+\Delta\sigma_{t_i}s_\theta(h_{t_i},\sigma_{t_i}),$$9

The reverse kernel is parameterized as

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$0

with

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$1

The dataset contains 13.25 hours of synchronized radar–respiration data from 12 healthy subjects, with 8 subjects for training and 4 for testing (Wang et al., 21 Mar 2026). Using only 20 reverse steps, the method reports state-of-the-art waveform and respiratory-rate performance: CS $$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$2, MSE $$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$3, MAE $$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$4 BPM, RMSE $$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$5 BPM, and SD $$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$6 BPM. Ablations show that removing self-attention, removing conditional phase tokens, or replacing banded cross-attention with naive concatenation all degrade performance. The architecture therefore uses diffusion not only as a denoiser, but as a temporally aligned observation-conditioned inverse model tuned to the physics of radar phase contamination by micromotions.

5. Diffuse scattering, diffuse multipath, and fading physics

Long before generative diffusion entered mmWave research, “diffusion” referred to diffuse scattering and diffuse multipath. RT-ICM provides a site-specific intra-cluster model in which one rough reflecting surface generates a cluster composed of a specular component and many diffuse rays (Yaman et al., 2019). The theoretical cluster CIR is

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$7

and diffuse power is shaped by the directive scattering pattern

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$8

together with a roughness loss

$$h_{t_{i-1}}=z_i+\rho_i\Phi^T(\Phi\Phi^T)^{-1}(y-\Phi z_i).$$9

At 60 GHz in a classroom, the model reports maximum cluster AoA error of approximately 1 degree, mean angle spread error of approximately 9 degrees, and RMS error for cluster peak power of approximately 2.2 dB (Yaman et al., 2019).

FDTD-based rough-slab analysis generalizes this by computing angle-resolved diffuse reflection and transmission patterns for walls, doors, and windows, then embedding those patterns into a ray tracer as secondary sources (Bakirtzis et al., 2020). The paper derives critical roughness thresholds such as

$-15$0

for reflection and shows that, at 28 GHz, diffuse reflection becomes significant beyond roughly $-15$1. In the reported indoor example, adding FDTD-derived diffuse rays improves RSS prediction in shadowed corridor and room regions that are underestimated by specular-only tracing (Bakirtzis et al., 2020).

At the small-scale fading level, “Better than Rician” argues that indoor directive mmWave links are often better modeled by Two-Wave with Diffuse Power rather than Rician fading (Zöchmann et al., 2018). The complex envelope is

$-15$2

with

$-15$3

Measurement campaigns at 60 GHz found that TWDP is favored over Rician when the steerable antenna points toward reflecting objects or is slightly misaligned at line-of-sight; the LoS $-15$4-factor was far above 20 dB, desk reflection had $-15$5 dB, and other reflections were around 10 dB (Zöchmann et al., 2018). The important point is that directive beams increase the relative weight of one or two dominant specular waves, but do not remove the diffuse floor.

Diffuse multipath can also be exploited rather than suppressed. “5G Positioning and Mapping with Diffuse Multipath” organizes MIMO-OFDM measurements into a multi-way tensor, performs CP decomposition, and applies tensor-ESPRIT to estimate continuous AoD, AoA, and delay parameters even in the absence of a specular component (Wen et al., 2019). The geometric channel model is

$-15$6

and the receiver position is obtained as a least-squares intersection of path-defined lines. In NLOS, the paper reports positioning RMSE improving from approximately 0.386 m at $-15$7 dB to approximately 0.0518 m at 10 dB for $-15$8 (Wen et al., 2019). This directly contradicts the common assumption that only specular mmWave paths are geometrically useful.

A further wave-physics usage appears in “Phase Singularity Diffusion,” which tracks the motion of singularities in transmitted speckle and shows that the mean-squared displacement grows linearly with frequency shift,

$-15$9

with the product of the phase-singularity diffusion coefficient and photon diffusion coefficient scaling with the square of effective sample length (Cheng et al., 2014). Although demonstrated in microwave experiments, the paper frames the method as scale-invariant, and this suggests a conceptual bridge between statistical speckle transport and mmWave diffuse propagation.

6. Practical implications, ambiguities, and future directions

Across learned and physical strands, several recurrent design principles are visible. Learned mmWave-Diffusion methods almost always use strong conditioning: DiffPace conditions on the linear measurement operator and HPSM beamspace; Radar-diffusion conditions on radar BEV; the range-image method conditions on radar range images; mmAnomaly conditions on RGBD-derived geometry and semantics; Diffusion$10$0 conditions on 3D point clouds, sparse RSSI, transmitter location, and frequency; and respiration sensing conditions on the observation itself (Hu et al., 21 Oct 2025, Luan et al., 2024, Wu et al., 4 Mar 2025, Toha et al., 1 Apr 2026, Park et al., 2 Oct 2025, Wang et al., 21 Mar 2026). This suggests that unconditional synthesis is rarely the objective in mmWave work; the dominant use case is conditional inverse modeling under strong physical constraints.

Inference efficiency is another unifying theme. Deterministic probability-flow ODE stepping in DiffPace, observation-anchored 20-step respiration denoising, one-step or accelerated latent diffusion variants in radar and anomaly pipelines, and 20-step DDIM-like heatmap generation all aim to reduce the prohibitive latency of long stochastic reverse chains (Hu et al., 21 Oct 2025, Wang et al., 21 Mar 2026, Wu et al., 4 Mar 2025, Park et al., 2 Oct 2025). A plausible implication is that practical mmWave-Diffusion systems will continue to prefer short deterministic samplers, plug-and-play corrections, or distilled surrogates over classical high-step DDPM inference.

The term also retains a non-generative meaning in vehicular networking. In “Deep-Reinforcement-Learning-Based Distributed Vehicle Position Controls for Coverage Expansion in mmWave V2X,” mmWave-Diffusion denotes the early-adoption phase in which only a fraction of vehicles carry mmWave radios (Taya et al., 2018). There, the technical problem is LOS relay formation under blockage, and the proposed A3C-based distributed position-control policy lets controllable vehicles make lane changes and longitudinal moves using only local observations. At $10$1, $10$2, and $10$3, the RL policy achieved about 1.7x the coverage of random movement (Taya et al., 2018). This usage is terminologically separate from diffusion models, but encyclopedically relevant because it appears explicitly under the same label.

Current limitations are correspondingly heterogeneous. Learned methods depend on paired supervision, synthetic-to-real transfer, calibration quality, and conditioning fidelity; physical diffuse-propagation models depend on roughness statistics, material parameters, and sometimes 2-D or first-order approximations. Future directions stated in the cited works include multi-user channel estimation, adaptive pilot and beamformer optimization, consistency-model acceleration, fast diffusion distillation, explicit radar physics in conditioning, OOD-aware context extraction, online learning for beam management, and full 3-D extensions of diffuse-scattering libraries (Hu et al., 21 Oct 2025, Toha et al., 1 Apr 2026, Azarbahram et al., 9 Apr 2026, Bakirtzis et al., 2020). Taken together, these lines of work indicate that mmWave-Diffusion is not a narrow subfield but a convergence zone between inverse problems, generative modeling, wave propagation, and geometry-aware sensing.