iRadioDiff: Physics-Informed Indoor Radio Maps

Updated 5 July 2026

The paper integrates material-aware prompts, multipath-critical priors, and boundary-weighted learning into a sampling-free diffusion framework for indoor radio map synthesis.
It employs a U-Net backbone with cross-attention to condition on physics prompts and indoor geometry, achieving lower RMSE and improved localization performance.
Empirical results on indoor layouts demonstrate that iRadioDiff preserves signal discontinuities, leading to enhanced received-signal-strength prediction and robust localization.

Searching arXiv for the primary iRadioDiff paper and closely related RadioDiff works to support the article. iRadioDiff is a physics-informed diffusion framework for indoor radio map construction and localization that generates indoor radio maps from geometry and materials alone rather than from in-situ measurements. It is formulated as a sampling-free diffusion-based framework for indoor radio map construction, conditioned on access point positions, a physics-informed prompt encoded by material reflection and transmission coefficients, and multipath-critical priors including diffraction points, strong transmission boundaries, and line-of-sight contours. The method is designed to model the heterogeneous and multipath-rich nature of indoor environments, where sharp, nonstationary field discontinuities arise at walls, doors, corners, and wedges, and it targets both high-fidelity radio map generation and received-signal-strength-based indoor localization (Wang et al., 25 Nov 2025).

1. Definition and scope

iRadioDiff addresses indoor radio map construction as a conditional generative modeling problem over a discretized two-dimensional region. The ground-truth radio map $P^*(x)$ records received signal strength over a grid $G$ for a known access point position $s$ , and the model outputs an estimate $\hat{P}$ covering the full grid (Wang et al., 25 Nov 2025). In the formulation reported for iRadioDiff, the conditioning set comprises four classes of inputs: an access point location encoded as a Gaussian heatmap, material-aware physics prompts $H_r(x)$ and $H_t(x)$ representing effective reflection and transmission coefficients, multipath-critical priors derived from diffraction-point candidates and strong-transmission boundaries, and a line-of-sight mask obtained by visibility analysis from the access point (Wang et al., 25 Nov 2025).

The framework is explicitly motivated by the mismatch between indoor propagation physics and conventional learning assumptions. Full-wave electromagnetic solvers are described as prohibitively expensive at room scale, while purely data-driven methods are characterized as oversmoothing discontinuities and often relying on sparse measurements or homogeneous-material assumptions that are misaligned with indoor propagation (Wang et al., 25 Nov 2025). Within the broader RadioDiff literature, this places iRadioDiff alongside diffusion-based radio-map models, but with a distinct emphasis on indoor geometry, material heterogeneity, and localization downstream of map generation. Related diffusion systems address radio map generation from sparse fragments or transmitter locations (Luo et al., 11 Jan 2025), Bayesian inverse reconstruction from noisy sparse measurements (Wang et al., 19 Apr 2025), Helmholtz-informed multipath-aware synthesis (Wang et al., 22 Apr 2025), and volumetric 3D radio map construction (Wang et al., 16 Jul 2025).

A recurrent source of confusion is nomenclature. Several earlier and contemporaneous papers do not use the name “iRadioDiff” even when their pipelines are diffusion-based and radio-domain specific. In the radio-frequency-fingerprinting literature, for example, the name is a later shorthand applied to a diffusion-denoising RFFI pipeline rather than an original method name (Yin et al., 7 Mar 2025). By contrast, “iRadioDiff” is the actual title of the indoor radio-map framework discussed here (Wang et al., 25 Nov 2025).

2. Diffusion formulation and generative mechanism

The generative backbone is a conditional, continuous-time diffusion model operating directly in the image space of radio maps, instantiated with a U-Net backbone that accepts time $t \in [0,1]$ and spatial conditioning through cross-attention (Wang et al., 25 Nov 2025). The paper places the method in the score-based diffusion family. In general form, the forward stochastic differential equation is

$d x = f(x,t)\,dt + g(t)\,d w,$

and the reverse-time SDE is

$d x = [f(x,t) - g(t)^2 \nabla_x \log p_t(x)]\,dt + g(t)\,d\bar{w},$

with the score approximated by a neural network conditioned on the physics prompts and priors (Wang et al., 25 Nov 2025).

iRadioDiff adopts a decoupled diffusion model inspired by RadioDiff. Under the canonical parametrization reported in the paper, the forward marginal is

$q(x_t \mid x_0) = N\!\left(x_0 + \int_0^t f_\tau d\tau,\; tI\right).$

This formulation separates deterministic contraction from stochastic perturbation, and the paper states that the reverse-time transition admits an analytical form that enables single-step or very-few-step generation. The resulting inference procedure is described as “sampling-free” in the sense that it avoids the long stochastic chains typical of DDPM-style samplers (Wang et al., 25 Nov 2025).

Training is performed by sampling $G$ 0, perturbing the ground-truth radio map $G$ 1 to $G$ 2, and training the U-Net to predict both the drift $G$ 3 and the added noise $G$ 4 with a mean-squared objective (Wang et al., 25 Nov 2025). The same paper also states a dataset-level supervised mapping objective,

$G$ 5

which situates the diffusion training within a broader conditional reconstruction setting (Wang et al., 25 Nov 2025).

This design differs from discrete DDPM radio-map generators such as RM-Gen, which use a standard variance-preserving forward process over $G$ 6 steps and ancestral reverse sampling (Luo et al., 11 Jan 2025), and from inverse-solver frameworks such as RadioDiff-Inverse, which use a diffusion prior inside Bayesian posterior sampling rather than direct conditional synthesis (Wang et al., 19 Apr 2025). It also differs from RMPrior, where acceleration is achieved by mid-start truncation from a matched propagation prior rather than by an intrinsically sampling-free generator (Guo et al., 2 Jun 2026). A plausible implication is that iRadioDiff trades some posterior-flexibility for direct inference speed and edge preservation, which is particularly aligned with indoor deployment and localization.

3. Physics-informed conditioning and multipath priors

The distinguishing feature of iRadioDiff is the form of its conditioning. The material-aware physics prompt is represented by co-registered spatial fields $G$ 7 and $G$ 8, corresponding to effective reflection and transmission coefficients derived from local material properties and propagation geometry (Wang et al., 25 Nov 2025). The paper does not impose a single explicit closed form for these maps, but it frames them as spatial prompts aligned with the floorplan grid and consumed through cross-attention in the U-Net (Wang et al., 25 Nov 2025).

The second major conditioning block consists of multipath-critical priors. The paper identifies three such priors: diffraction points, strong transmission boundaries, and line-of-sight contours (Wang et al., 25 Nov 2025). Candidate diffraction points are extracted from the reflection map under a local neighborhood rule: the candidate cell has non-zero reflectance, at least one of its four neighbors has non-zero reflectance, and at least two neighbors have zero reflectance, indicating a convex reflectance discontinuity (Wang et al., 25 Nov 2025). These candidates are then subject to several pruning stages. Radial connections that directly link the access point to a candidate are removed, candidates are discarded if the access point and the candidate’s outward normal lie in the same directional quadrant, and corners in the same room as the access point may also be removed if their effect on RSS discontinuity is negligible (Wang et al., 25 Nov 2025).

Strong transmission boundaries are extracted by thresholding the transmission map $G$ 9 and connecting adjacent cells into boundary segments (Wang et al., 25 Nov 2025). The line-of-sight mask $s$ 0 is computed by rotational scanning or ray casting from the access point, producing both a visibility field and a LoS–NLoS transition contour (Wang et al., 25 Nov 2025). These structures are then channelized and injected as spatial conditioning maps alongside the access point heatmap and the material prompts (Wang et al., 25 Nov 2025).

The broader research context suggests why such priors matter. RadioDiff- $s$ 1 argues that multipath-critical structures correspond to regions associated with negative wave numbers in a Helmholtz-based analysis and uses a dual-diffusion architecture to infer singularities before reconstructing the full radio map (Wang et al., 22 Apr 2025). RadioDiff-Loc similarly exploits the physical informativeness of obstacle vertices for non-line-of-sight localization with sparse radio map estimation (Wang et al., 2 Sep 2025). iRadioDiff does not reuse those exact mechanisms, but they reinforce the same principle: explicitly marking discontinuity-bearing structures improves generative fidelity where conventional convolutional smoothness priors are weakest.

4. Boundary-weighted learning and field discontinuities

iRadioDiff places particular emphasis on “boundary-weighted objectives” to preserve sharp transitions in the electromagnetic field (Wang et al., 25 Nov 2025). The paper states that the diffusion training uses boundary-weighted learning signals emphasizing regions near the multipath-critical prior contours and line-of-sight boundaries. While an explicit closed-form weighting function is not printed in the paper, the description is clear that the weights are constructed from the prior masks and are larger near discontinuities (Wang et al., 25 Nov 2025).

This matters because indoor radio maps violate spatial stationarity in a structured way. Reflection, penetration, and diffraction induce abrupt transitions at wall boundaries, doors, and corners, while line-of-sight occlusion creates shadow fronts that standard smooth regressors tend to blur (Wang et al., 25 Nov 2025). The boundary-weighted objective is intended to shift optimization pressure toward these regions, counteracting the smoothing bias of convolutional architectures and improving downstream localization sensitivity (Wang et al., 25 Nov 2025).

A plausible interpretation is that iRadioDiff’s performance gains do not arise solely from adding more channels, but from aligning the optimization target with the radio-map error modes that most strongly affect localization. This interpretation is consistent with the ablation reported in the paper: removing all physics-related priors, including $s$ 2, $s$ 3, $s$ 4, and $s$ 5, substantially degrades RMSE, PSNR, LPIPS, FID, and localization error (Wang et al., 25 Nov 2025).

5. Data, protocols, and empirical performance

The experiments use a subset of the Indoor Radio Map Dataset at 3.5 GHz and a spatial resolution of $s$ 6 m (Wang et al., 25 Nov 2025). The ray-traced ground truth is generated with up to 8 reflections, 10 transmissions, and 2 diffractions, with transmitter and receiver height both fixed at $s$ 7 m and antenna pattern differences ignored (Wang et al., 25 Nov 2025). Two evaluation protocols are defined across 25 layouts. In antenna location generalization (ALG), all 25 layouts are used for training, with 40 of 50 access points per layout for training and 10 for testing. In zero-shot layout generalization (ZLG), 20 layouts with 50 access points each are used for training and 5 unseen layouts with 50 access points each are reserved for testing (Wang et al., 25 Nov 2025).

The paper compares iRadioDiff against RadioUNet, RME-GAN, and SIP2Net using RMSE, PSNR, LPIPS, and FID for radio-map construction, and mean localization error in meters for RSS-based localization (Wang et al., 25 Nov 2025). The main reported results are summarized below.

Protocol	Method	Key reported results
ALG	iRadioDiff	RMSE 6.357, PSNR 32.24, LPIPS 0.2742, FID 145.2, localization 7.860 m
ALG	RadioUNet	RMSE 9.349, PSNR 28.92, LPIPS 0.3667, FID 240.12, localization 12.09 m
ALG	SIP2Net	RMSE 24.55, PSNR 21.24, LPIPS 0.3121, FID 196.8, localization 21.41 m
ALG	RME-GAN	RMSE 103.1, PSNR 7.902, LPIPS 0.5459, FID 308.1, localization 29.60 m
ZLG	iRadioDiff	RMSE 7.010, PSNR 31.45, LPIPS 0.3301, FID 192.6, localization 8.530 m
ZLG	RadioUNet	RMSE 7.868, PSNR 30.52, LPIPS 0.4188, FID 309.1, localization 12.16 m
ZLG	SIP2Net	RMSE 47.43, PSNR 15.32, LPIPS 0.3289, FID 197.2, localization 25.40 m
ZLG	RME-GAN	RMSE 93.83, PSNR 8.723, LPIPS 0.5236, FID 270.5, localization 19.33 m

These results show that iRadioDiff is the only reported method remaining below 10 m mean localization error in both protocols (Wang et al., 25 Nov 2025). The ablation without physics-related priors further supports the architecture’s design rationale. In ALG, removing the priors degrades RMSE from 6.357 to 9.619, PSNR from 32.24 to 28.65, LPIPS from 0.2742 to 0.6260, FID from 145.2 to 351.1, and localization error from 7.860 m to 8.553 m (Wang et al., 25 Nov 2025). In ZLG, the same ablation worsens RMSE from 7.010 to 11.75, PSNR from 31.45 to 27.06, LPIPS from 0.3301 to 0.3717, and FID from 192.6 to 242.1, with localization also degrading (Wang et al., 25 Nov 2025).

These evaluations position iRadioDiff relative to the broader radio-map literature. RM-Gen reports strong conditional radio-map generation from sparse fragments and transmitter coordinates in outdoor and indoor environments, but its task and metrics are framed around error-tolerance-rate accuracy rather than localization (Luo et al., 11 Jan 2025). RadioDiff-Inverse and RadioDiff-Inv2 focus on inverse estimation under sparse noisy measurements, including environmental perception and location drift (Wang et al., 19 Apr 2025, Wang et al., 7 Jun 2026). RadioDiff-3D extends diffusion-based radio-map construction to volumetric pathloss, DoA, and ToA fields (Wang et al., 16 Jul 2025). iRadioDiff’s specific empirical niche is measurement-free indoor radio-map synthesis with localization as the target downstream task (Wang et al., 25 Nov 2025).

6. Localization pipeline, interpretation, and limitations

Localization in iRadioDiff is performed by fingerprint matching using K-nearest neighbors with $s$ 8 on RSS vectors across access points (Wang et al., 25 Nov 2025). Given an observed RSS vector $s$ 9, the estimate is

$\hat{P}$ 0

where $\hat{P}$ 1 denotes the radio-map fingerprint at candidate position $\hat{P}$ 2 (Wang et al., 25 Nov 2025). In single-access-point settings this reduces to scalar matching against the generated intensity field, while in multi-access-point settings it becomes a standard multi-dimensional fingerprinting problem (Wang et al., 25 Nov 2025).

The method assumes a 2D planar environment, quasi-static materials, a single carrier at 3.5 GHz, equal transmitter and receiver heights, and no explicit antenna pattern modeling (Wang et al., 25 Nov 2025). These assumptions delimit its domain of validity. The paper notes potential failure modes in highly metallic or waveguiding environments, dynamic blockage scenarios, and settings with unresolved micro-geometry below the grid scale (Wang et al., 25 Nov 2025). It also points to extensions such as multi-AP and 3D or multi-floor formulations, dynamic environments, mmWave propagation with beam patterns and polarization, and richer diffraction priors (Wang et al., 25 Nov 2025).

That forward-looking perspective is consistent with subsequent work in the same family. RadioDiff-3D introduces a 3D $\hat{P}$ 33D benchmark with pathloss, DoA, and ToA over 20 receiver heights and supports both transmitter-known and sparse-observation settings (Wang et al., 16 Jul 2025). RadioDiff-Loc applies conditional diffusion to NLoS localization with power-invariant sparse radio-map estimation (Wang et al., 2 Sep 2025). RMPrior shows that matched propagation priors can be injected at intermediate diffusion timesteps to improve speed and fidelity in radio-map refresh tasks (Guo et al., 2 Jun 2026). These developments suggest that iRadioDiff may be understood ոչ merely as a single model, but as an indoor, physics-informed anchor point in a growing family of diffusion-based electromagnetic field constructors.

From an encyclopedic standpoint, the defining contribution of iRadioDiff is the integration of three elements into a single indoor RM pipeline: material-aware prompting, explicit multipath-critical priors, and boundary-weighted diffusion learning, all coupled to a sampling-free generative mechanism and validated by downstream RSS localization (Wang et al., 25 Nov 2025). This suggests a broader methodological lesson: in radio-map generation, the decisive improvement comes not simply from adopting diffusion, but from coupling diffusion to the discontinuity structure imposed by electromagnetic propagation.