Pixel-Level Residual Diffusion Transformer: Scalable 3D CT Volume Generation

A new generative framework called Pixel-Level Residual Diffusion Transformer (PRDiT) has been proposed to synthesize high-resolution 3D CT medical volumes directly at the voxel level, addressing long-standing computational and optimization challenges in the field [1]. The framework, detailed in a paper submitted to arXiv on June 18, 2026, introduces a two-stage training architecture designed to simplify optimization and improve training stability [1]. The first stage uses a local denoiser, an MLP-based blind estimator that operates on overlapping 3D patches to separate low-frequency structures efficiently. The second stage employs a global residual diffusion transformer with memory-efficient attention to model and refine high-frequency residuals across the entire volume [1]. This coarse-to-fine strategy avoids the limitations of an autoencoder bottleneck, a common constraint in previous models [1]. In experiments on the LIDC-IDRI and RAD-ChestCT datasets, PRDiT consistently outperformed state-of-the-art generative models, including HA-GAN, 3D LDM, and WDM-3D [1]. The performance was measured using 3D Fréchet Inception Distance (FID), Maximum Mean Discrepancy (MMD), and Wasserstein distance scores, with PRDiT achieving significantly lower values across all metrics [1]. The ability to generate high-fidelity 3D medical volumes has implications for data augmentation in diagnostic imaging, where the availability of large, annotated datasets often limits the development of robust machine learning models [1]. Generative models for medical imaging have historically struggled with the high dimensionality of 3D data, which demands substantial computational resources and can lead to unstable training dynamics [1]. The PRDiT architecture addresses this by decoupling the generation of low-frequency structure from high-frequency detail, a technique that has parallels in other domains where coarse-to-fine modeling has proven effective for complex data distributions [1]. The use of a residual diffusion transformer for global refinement allows the model to scale to entire volumes without the memory bottlenecks that typically constrain transformer-based architectures on 3D data [1].