ST-CGMDiff: Sparse-Conditioned Cross-Frequency Spatio-Temporal Latent Diffusion for Dynamic 3D Channel Gain Map Reconstruction

arXiv:2611.06220 · eess.SP · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "ST-CGMDiff: Sparse-Conditioned Cross-Frequency Spatio-Temporal Latent Diffusion for Dynamic 3D Channel Gain Map Reconstruction".

Dev: Here's the gist of the paper. Accurate and efficient reconstruction of channel gain maps (CGMs) is essential for environment-aware wireless communications This paper proposes ST-CGMDiff, a spatio-temporal latent diffusion framework that integrates sparse-observation conditioning and cross-frequency modeling to reconstruct dynamic 3D CGMs across multiple carrier frequencies and varying Rx heights within a unified model Problem Formulation The objective is to reconstruct the 3D CGMs over a considered region at different time slots and carrier frequency bands The channel gain between a Tx–Rx pair is described by gt: R cubed × R cubed × F → R, where gt (ptx, prx; f) denotes the channel gain between the Tx at ptx and the Rx at prx The reconstruction problem focuses on recovering the channel gains of unobserved grid cells from available spatial, frequency, and temporal information The target is to estimate Xt from the record O t f = (Ys,f,Ms,f)s=0,…,t, f∈F ST-CGMDiff Architecture The proposed ST-CGMDiff framework reconstructs dynamic 3D CGMs through three main processing stages: Spatial Condition Encoding, Frequency-Conditioned Temporal Fusion, and Height-Aware Latent Diffusion and Decoding First, the current sparse CGM observation, observation mask, Tx and Rx z-coordinates, and carrier frequency are jointly encoded to obtain the current frequency-aware spatial condition Second, previously reconstructed CGMs are encoded into the latent space and processed by a frequencyconditioned ConvGRU to extract causal temporal information Third, the resulting fused condition then guides a height-aware latent diffusion network, in which BiConvLSTM modules exchange information across the ordered Rx z-layers Key Architectural Components The framework incorporates several key components to achieve its goals: 1.

Rosa: Spatial Condition Encoding involves constructing the current spatial condition tensor Ct f = Cat

Yt,f,Mt,f, Az,rx, Az,tx, Af: .

Dev: The subpixel condition encoder uses pixel unshuffle to reduce the horizontal spatial resolution of the sparse observation and signed observation mask while preserving their sampling locations.

Title and authors: Rosa: The title of the paper, ST-CGMDiff: Sparse-Conditioned Cross-Frequency Spatio-Temporal Latent Diffusion for Dynamic three dee Channel Gain Map Reconstruction, really tells you what it’s trying to achieve <ref:2611.06220#pg1,ST-CGMDiff: Sparse-Conditioned Cross-Frequency Spatio-Temporal Latent Diffusion for Dynamic>.

Dev: It’s not just about static maps at one frequency anymore; it explicitly mentions dynamic maps across multiple frequencies and time slots, which is where the real challenge in wireless environments lies.

Rosa: Right. And the core concept is this latent diffusion model that integrates sparse observation conditioning and cross-frequency modeling to reconstruct those three dimensional channel gain maps across different carrier frequencies and varying receiver heights within a single model structure.

Taro: So, it’s unifying things that used to require separate models for each frequency or for each height level, which is a big simplification if it works.

Dev: It sounds like they are tackling the difficulty of dynamic multifrequency environments where both the spatial and temporal variations are happening at once. What about the authors?

Rosa: The authors are Tianxiang Li, Yilong Chen, Han Hu, Yong Zeng, and Jie Xu. They’re folks with expertise in this area who developed this unified framework.

Taro: I wonder what their background is in terms of diffusion models versus traditional channel estimation methods like ray tracing or Kriging that we see elsewhere.

Dev: Ray tracing needs detailed environmental descriptions, and measurement-driven methods recover values directly from measurements without needing that explicit geometry, which is a key comparison point for this work.

The paper's summary: Rosa: So to summarize the paper’s main points, ST-CGMDiff reconstructs dynamic three dimensional channel gain maps over a region at different time slots and carrier frequencies using limited measurements.

Dev: The framework involves jointly encoding the input parameters like sparse observations, observation masks, Tx and receiver z coordinates, and carrier frequency to get a current frequency aware spatial condition.

Taro: Then they take previously reconstructed maps and process them through a frequency conditioned convolutional recurrent neural network to pull out causal temporal information.

Rosa: And that fused condition then guides a height aware latent diffusion network where biconvolutional lstm modules exchange information across the ordered receiver z layers.

Dev: That sounds like a sophisticated way to handle the spatial structure in three dimensions without assuming everything is perfectly spaced in terms of those z coordinates.

Taro: So, they are using the temporal modeling to capture history and the cross layer exchange to handle that vertical dimension flexibly. It’s about building a robust system for these complex channel variations over time.

Rosa: Right, it’s about fusing all these elements—spatial condition encoding, temporal fusion, and height aware latent diffusion—into one cohesive framework for reconstruction.

The paper's improvements: Dev: The paper highlights several key architectural components that they built to achieve this. First, the spatial condition encoding uses pixel unshuffle to reduce the horizontal spatial resolution of sparse observations and the observation mask while keeping their sampling locations intact.

Taro: That reduction in horizontal resolution is a smart move if it helps manage the complexity of those sparse inputs before feeding them into a larger model.

Rosa: Next, they have frequency conditioned temporal fusion using stacked causal convolutional recurrent neural networks to process historical latent representations, which are then processed by a frequency conditioning network to produce modulation parameters alpha and beta for feature wise modulation.

Dev: So they’re not just doing standard sequence modeling; they’re adapting the temporal processing based on the specific carrier frequency being used. That helps capture how different frequencies behave differently over time.

Taro: And then there’s the height aware latent diffusion part where those biconvolutional lstm modules exchange information across ordered receiver z layers, which lets them process different numbers and selections of receiver z layers without assuming uniformly spaced z coordinates.

Rosa: Finally, they perform reconstruction through few step diffusion in a compact latent space followed by full resolution residual refinement to get the final channel gain map with improved spatial fidelity.

Dev: That few step diffusion part sounds like it significantly cuts down on the computational cost of the reconstruction loop itself, which is good for real-time considerations.

Conclusion: Rosa: So, wrapping up, ST-CGMDiff offers a unified frequency conditioned spatio-temporal latent diffusion framework for dynamic three dimensional channel gain map reconstruction from sparse measurements.

Dev: By jointly exploiting temporal correlations, cross z layer dependencies, and frequency information, this framework enables efficient reconstruction across multiple carrier frequencies and flexible receiver heights without needing separate models for each band or height.

Taro: The paper shows that by using one jointly trained checkpoint across the four frequency bands instead of separate ones for each band, they maintain stable completion performance even at unseen frequencies like four point seven gigahertz and fifteen point one gigahertz without retraining or fine-tuning.

Rosa: And they also showed it reduces the number of reverse diffusion steps from fifty to ten and cuts the average inference latency from seventeen point seven seconds down to just zero point eight four four seconds.

Dev: That latency reduction is substantial, meaning this could be viable for more demanding real-time applications where you need fast channel updates.

Taro: It’s impressive that they maintained consistent performance at those unseen frequencies, suggesting the interpolation capability of the frequency conditioned reconstruction framework is quite robust.

Rosa: It really moves the field forward by providing a flexible and efficient way to estimate these complex three dee channel maps dynamically in real-world wireless systems <ref:2611.06220#pg1>.

Tianxiang Li, Yilong Chen, Han Hu, Yong Zeng, Jie Xu

Chien-Shiung Wu College, Southeast University · School of Science and Engineering and the Future Network of Intelligence Institute, The Chinese University of Hong Kong (Shenzhen) · Beijing Institute of Technology · National Mobile Communications Research Laboratory · Purple Mountain Laboratories

eess.SP

Submitted: 2026-11-04

Updated: 2026-11-04

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 95/100

The gist: Accurate and efficient reconstruction of channel gain maps (CGMs) is essential for environment-aware wireless communications This paper proposes ST-CGMDiff…

Key concepts

3D CGM Reconstruction
This involves estimating the channel gain between a transmitter (Tx) and receiver (Rx) across three dimensions: spatial location (x, y), frequency bands, and Rx height. The goal is to recover the full 3D map of these gains from limited, sparse measurements.
Latent Diffusion Framework
This technique uses a generative model approach where the reconstruction process is framed as reversing a diffusion process. Instead of directly predicting the complex map, the model learns to iteratively refine a compact latent representation, which is then decoded into high-resolution channel gain maps.

Terminology

Summary

Accurate and efficient reconstruction of channel gain maps (CGMs) is essential for environment-aware wireless communications This paper proposes ST-CGMDiff, a spatio-temporal latent diffusion framework that integrates sparse-observation conditioning and cross-frequency modeling to reconstruct dynamic 3D CGMs across multiple carrier frequencies and varying Rx heights within a unified model

Problem Formulation

The objective is to reconstruct the 3D CGMs over a considered region at different time slots and carrier frequency bands The channel gain between a Tx–Rx pair is described by gt: R cubed × R cubed × F → R, where gt (ptx, prx; f) denotes the channel gain between the Tx at ptx and the Rx at prx The reconstruction problem focuses on recovering the channel gains of unobserved grid cells from available spatial, frequency, and temporal information The target is to estimate Xt from the record O t f = (Ys,f,Ms,f)s=0,…,t, f∈F

ST-CGMDiff Architecture

The proposed ST-CGMDiff framework reconstructs dynamic 3D CGMs through three main processing stages: Spatial Condition Encoding, Frequency-Conditioned Temporal Fusion, and Height-Aware Latent Diffusion and Decoding First, the current sparse CGM observation, observation mask, Tx and Rx z-coordinates, and carrier frequency are jointly encoded to obtain the current frequency-aware spatial condition Second, previously reconstructed CGMs are encoded into the latent space and processed by a frequencyconditioned ConvGRU to extract causal temporal information Third, the resulting fused condition then guides a height-aware latent diffusion network, in which BiConvLSTM modules exchange information across the ordered Rx z-layers

Key Architectural Components

The framework incorporates several key components to achieve its goals:

  1. Spatial Condition Encoding involves constructing the current spatial condition tensor Ct f = Cat [Yt,f,Mt,f, Az,rx, Az,tx, Af]. p. 4 The subpixel condition encoder uses pixel unshuffle to reduce the horizontal spatial resolution of the sparse observation and signed observation mask while preserving their sampling locations. p. 4

  2. Frequency-Conditioned Temporal Fusion utilizes a stacked causal ConvGRU [40] to process historical latent representations, which are then processed by a frequency-conditioning network to produce modulation parameters αf and βf for feature-wise modulation. p. 5

  3. Height-Aware Latent Diffusion and Decoding employs BiConvLSTM modules to exchange information across the ordered Rx z-layers, allowing the model to process different numbers and selections of Rx z-layers without assuming uniformly spaced z-coordinates Finally, reconstruction is performed through few-step diffusion in a compact latent space, followed by full-resolution residual refinement

Training and Loss Functions

To train the proposed ST-CGMDiff framework, the authors jointly optimize three objectives:

** Latent-space training objective:**

The complete latent objective is defined as Llat = Latt + Lnoise + λcleanLclean. This includes the attenuation prediction loss (Latt), noise prediction loss (Lnoise), and clean latent reconstruction loss (Lclean)

** Pixel-domain objective:**

The pixel-domain objective is defined as Lpix = Lmiss + λgradLgrad + λmsLms + λssimLssim. This loss includes the missing-region reconstruction loss (Lmiss), spatial-gradient loss (Lgrad), multiscale reconstruction loss (Lms), and structural similarity loss (Lssim)

** Temporal-prediction objective:**

The temporal prediction head is supervised independently with the objective Lpred = omegat,f ⊙ Xb temp t,f − Xt,f The complete training objective is L = Llat + λpixLpix + λpredLpred

Experimental Results and Analysis

Experiments on a dynamic multifrequency dataset demonstrate that the proposed ST-CGMDiff consistently outperforms diffusion-based and conventional reconstruction baselines in reconstructing high-resolution CGMs from sparse observations The results show that the jointly trained ST-CGMDiff achieves the lowest average RMSE and NMSE and the highest PSNR under the evaluated protocols at the four training frequencies Furthermore, neither unseen frequency exhibits an obvious degradation relative to the neighboring training frequencies, supporting interpolation capability The proposed method also reduces the number of reverse diffusion steps from 50 to 10 and decreases the average inference latency from 17.779 s to 0.844 s

Ablation Study Findings

The ablation study demonstrated the complementary contributions of temporal modeling, information exchange across Rx z-layers, and high-resolution refinement Removing the temporal ConvGRU, z-wise BiConvLSTM, and high-resolution refiner reduces PSNR by 1.669, 2.827, and 5.241 dB, respectively The substantial performance degradation without the refiner highlights the importance of preserving fine spatial details during latent reconstruction

Conclusion

This paper proposed ST-CGMDiff, a unified frequencyconditioned spatio-temporal latent diffusion framework for dynamic 3D CGM reconstruction from sparse measurements By jointly exploiting temporal correlations, cross-z-layer dependencies, and frequency information, the proposed framework enables efficient reconstruction across multiple carrier frequencies and flexible Rx heights The framework also generalized to unseen carrier frequencies without retraining and maintained consistent performance across different sampling budgets For contiguous-region completion, the frequency-conditioned configuration improved PSNR by 1.480 dB over 3DCKMDiff while reducing inference latency by approximately 95. p. 83% The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed ST-CGMDiff is also tested at unseen 4. p. 87 and 15.1 GHz without retraining or fine-tuning The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11. p. 92% and 14.2%, increases SSIM from 0. p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 978009 to 0.80789, and improves PSNR by 1. p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces the number of reverse diffusion steps from 50 to 10 and decreases the average inference latency from 17.779 s to 0.844 s The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11. p. 92% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11. p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 9 p. 92% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework The proposed method reduces RMSE and NMSE by approximately 11.2% and 14.2%, increases SSIM from 0.78009 to 0.80789, and improves PSNR by 1.480 dB The proposed method uses one jointly trained checkpoint across the four frequency bands and explicitly conditions on carrier frequency, whereas 3DCKMDiff uses a separate checkpoint for each band The proposed method also maintains stable completion performance at the two frequencies unseen during training The resulting performance remains consistent with that of the neighboring seen frequency bands, further supporting the interpolation capability of the frequency-conditioned reconstruction framework <ref:2611. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved system can reconstruct dynamic 3D channel gain maps across multiple frequency bands and varying Rx heights using a single trained model, eliminating the need for separate models for different frequencies or explicit environmental maps. This enables a unified model to accommodate multiple carrier frequencies, varying Tx z-coordinates, and flexible Rx z-layer selections without requiring separate models.

  2. Improved system can exploit temporal correlations through a frequency-conditioned ConvGRU and capture ordered dependencies among selected Rx zlayers using BiConvLSTM modules to extract causal information from previous model reconstructions. This allows the system to maintain closed-loop reconstruction over consecutive time slots by adaptively fusing historical information with current sparse observations.

  3. Improved system can achieve high spatial fidelity in reconstructed channel gain maps by employing a few-step latent diffusion scheme followed by high-resolution residual refinement, which reduces the iterative reconstruction burden while recovering fine spatial details. This is achieved through the formulation of a few-step latent diffusion followed by high-resolution residual refinement to generate the final CGM with improved spatial fidelity.

  4. Improved system can perform accurate contiguous-region completion by utilizing a spatially adaptive fusion mechanism that balances historical information and current observations, as described in the paper's contributions. This mechanism is designed to enable accurate contiguous-region completion with substantially lower inference cost.

Sources

Related papers