⚡ electrical engineering

CDiT: Conditional Diffusion Transformer for Geometry-Aware Terahertz Cross Far- and Near-Field Channel Generation

This paper proposes the Conditional Diffusion Transformer (CDiT), a geometry-aware generative framework that leverages a hybrid planar-spherical wave model and transformer-based diffusion to achieve high-fidelity, controllable synthesis of complex Terahertz cross far- and near-field channels, significantly outperforming existing model-based approaches.

Original authors: Zhengdong Hu, Chong Han

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhengdong Hu, Chong Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a perfect map of a city, but the city is made of invisible sound waves (Terahertz waves) that behave very strangely. These waves are so high-frequency that they act like light, but they also have a unique quirk: depending on how far you are from the source, they can behave like flat sheets of water (far-field) or like ripples spreading out from a stone dropped in a pond (near-field).

Designing the next generation of super-fast wireless internet (6G and beyond) requires understanding these waves perfectly. But creating a computer model that simulates them is incredibly hard. It's like trying to predict the exact pattern of every single ripple in a massive ocean, where the rules change depending on where you stand.

Here is how the authors of this paper solved that problem, explained simply:

The Problem: The "One-Size-Fits-None" Map

Traditional ways of modeling these signals are like using a flat, 2D map to navigate a mountain range. They make too many simple guesses (like assuming all waves are flat sheets) and miss the complex, 3D reality of how signals bounce and curve in real life. Other methods try to calculate every single physics equation from scratch, but that takes so much computer power it's like trying to count every grain of sand on a beach to build a sandcastle.

The Solution: A "Smart Art Teacher" (CDiT)

The authors created a new tool called CDiT (Conditional Diffusion Transformer). Think of it as a highly skilled art teacher who learns to paint perfect landscapes by studying thousands of real photos.

Here is how the "teacher" works, broken down into three parts:

1. The "Noise-to-Picture" Process (Diffusion)
Imagine you have a beautiful, clear photo of a city (the real signal).

The Forward Process: The teacher starts by slowly adding static noise to the photo until it becomes a blurry, unrecognizable mess of gray pixels.
The Reverse Process: The teacher then learns to reverse this. Starting with a blank, noisy canvas, they learn to slowly remove the noise, step-by-step, to reveal a perfect, clear picture of the city again.
The Magic: Instead of just guessing, the teacher learns the rules of how the noise turns into a picture. This is more stable and reliable than older methods (like GANs) that often get confused and produce blurry or weird results.

2. The "Global Eye" (Transformer)
Older AI models look at a picture like a person looking through a tiny straw—they only see a small patch at a time. This is fine for simple things, but for Terahertz signals, the whole picture matters. A ripple on the left side of the ocean affects the right side.
The authors used a Transformer, which is like giving the teacher a giant, wide-angle lens. They can see the entire image at once. This allows them to understand long-distance connections and complex patterns that other models miss.

3. The "GPS Guide" (Conditional)
The most important part is that this teacher doesn't just paint random pictures. They are given a GPS coordinate (the position of the user).

If you tell the teacher, "I am standing 10 meters away," it paints a picture where the waves look like ripples (near-field).
If you say, "I am 1 kilometer away," it paints a picture where the waves look like flat sheets (far-field).
This makes the model "geometry-aware." It knows exactly how the signal should look based on where you are standing.

The Results: A Perfect Match

The authors tested their "teacher" against real-world data and other AI models:

Better than the competition: Their model produced signal maps that were 70% more accurate (in terms of structural similarity) than the previous best methods.
It gets the physics right: The generated signals didn't just look pretty; they matched the real physics of how energy spreads out in space. If you looked at where the signal energy was strongest, the AI's map matched the real-world map almost perfectly.

In a Nutshell

The paper presents a new way to simulate the complex, invisible world of Terahertz wireless signals. By combining a "noise-removing" learning process with a "wide-angle" AI brain and a "GPS" guide, they created a system that can generate highly realistic, location-specific signal maps. This helps engineers design better, faster, and more reliable wireless systems for the future without needing to build expensive physical prototypes for every single scenario.

Technical Summary: CDiT: Conditional Diffusion Transformer for Geometry-Aware Terahertz Cross Far- and Near-Field Channel Generation

Problem Statement
Accurate channel modeling is critical for the design and evaluation of Terahertz (THz) ultra-massive multiple-input multiple-output (UM-MIMO) systems. However, existing approaches face significant limitations. Traditional model-based methods often rely on simplified assumptions, such as sparsity or predefined parametric structures, which fail to capture the complex spatial variations and the coexistence of near-field and far-field propagation characteristics inherent in practical THz environments. While deep learning (DL) offers a data-driven alternative, existing generative models like Generative Adversarial Networks (GANs) suffer from training instability and mode collapse. Furthermore, recent diffusion models often rely on Convolutional Neural Networks (CNNs), which are limited by local receptive fields and struggle to model the long-range spatial dependencies and high-dimensional structures of THz UM-MIMO channels. Additionally, most existing generative frameworks focus exclusively on far-field scenarios, neglecting the hybrid near-/far-field effects introduced by extremely large antenna apertures.

Methodology
The paper proposes a Conditional Diffusion Transformer (CDiT) framework to address these challenges. The methodology is built upon three core components:

Hybrid Planar-Spherical Wave Model (HPSM) and Beamspace Representation:
To balance physical fidelity with computational tractability, the authors utilize the HPSM. This model leverages a widely-spaced multi-subarray architecture where intra-subarray propagation is modeled as planar waves, while inter-subarray propagation accounts for spherical wave effects. The channel is transformed into a sparse beamspace domain representation ( $H_b$ ). This transformation reduces the dimensionality of the generative task and preserves the physically interpretable angular power distribution, allowing the model to focus on dominant propagation clusters rather than dense, noisy coefficients.
Conditional Diffusion Process:
The channel modeling task is formulated as a conditional generative learning problem. The model learns to transform a latent Gaussian variable into a beamspace channel realization conditioned on a geometry-aware vector $p$ . This vector includes Euclidean distance, relative Cartesian coordinates, and angular information (encoded via sine/cosine to ensure continuity). The process follows a forward diffusion step that progressively adds noise to the channel data and a reverse denoising step that reconstructs the clean channel from noise, guided by the geometry condition.
Diffusion Transformer (DiT) Architecture:
Instead of using CNNs, the denoising backbone employs a Diffusion Transformer. The noisy channel input is patchified and embedded into a sequence of tokens. A specialized conditioning mechanism injects both the diffusion timestep and the user position information into the transformer backbone via Adaptive Layer Normalization (adaLN). The Multi-Head Self-Attention (MHSA) mechanism allows the model to capture global spatial dependencies across the entire high-dimensional channel representation, effectively modeling the complex correlations arising from hybrid-field propagation and ultra-large antenna apertures.

Key Contributions
The paper outlines three primary contributions:

Unified Hybrid-Field Generative Modeling: The development of a diffusion-based framework capable of unifying near-field, far-field, and cross-field propagation behaviors within a single generative model under the HPSM.
Geometry-Aware Transformer-Based Generation: The design of a DiT-based generator that incorporates position information as a conditioning signal. This architecture leverages global self-attention to capture long-range spatial dependencies, enabling controllable generation of channel realizations that preserve dominant propagation clusters and angular power distributions.
Experimental Validation: Comprehensive evaluation demonstrating that the proposed framework achieves stable convergence and significantly outperforms benchmark methods in terms of structural similarity and physical fidelity.

Results
Extensive experiments were conducted using a dataset of 50,000 THz channel realizations generated via QuaDRiGa, calibrated with real measurement statistics from indoor corridor scenarios at 306–321 GHz. The system utilized $N_t=256$ transmit and $N_r=64$ receive antennas.

Convergence: The framework demonstrated stable convergence, with testing loss stabilizing after approximately 60 epochs.
Structural Similarity: The proposed CDiT achieved an average Structural Similarity Index Measure (SSIM) of 0.75. This represents a significant improvement over a GAN-based baseline (0.43) and a CNN-based diffusion model (0.46), corresponding to a ~74.4% improvement over the GAN approach.
Power Distribution: The generated channels closely aligned with ground-truth data regarding angular power distributions in the beamspace domain. The model successfully reconstructed principal angular energy concentrations and spatial locations, confirming its ability to preserve physically meaningful spatial characteristics.

Significance and Claims
The authors claim that the proposed CDiT framework provides a promising data-driven paradigm for THz channel modeling in next-generation wireless systems. By combining the distribution modeling strength of diffusion models with the global dependency modeling of transformers, the framework overcomes the limitations of conventional planar-wave assumptions and local-receptive-field architectures. The paper asserts that this approach effectively preserves both the statistical and physically meaningful spatial characteristics of practical THz propagation environments, offering a robust solution for the design and evaluation of THz UM-MIMO systems where accurate channel characterization is paramount. The work does not claim to solve all THz challenges but specifically addresses the gap in high-fidelity, geometry-aware generative modeling for hybrid near-/far-field scenarios.

The Problem: The "One-Size-Fits-None" Map

The Solution: A "Smart Art Teacher" (CDiT)

The Results: A Perfect Match

In a Nutshell

More like this