How PRDiT Achieves Pixel-Perfect 3D CT Volumes Without the Usual Tradeoffs

A team from the University of Melbourne and Monash University has introduced PRDiT, a Pixel-Level Residual Diffusion Transformer designed to generate high-resolution 3D CT volumes directly at the voxel level. The model addresses the computational demands and optimization hurdles typical of generative models by adopting a coarse-to-fine strategy. Its two-stage architecture first uses an MLP-based blind estimator on overlapping 3D patches to separate low-frequency structures, then a residual diffusion transformer with memory-efficient attention refines high-frequency residuals across the entire volume. This approach avoids the limitations of an autoencoder bottleneck, preserving subtle anatomical details.

Extensive experiments on the LIDC-IDRI and RAD-ChestCT datasets show PRDiT consistently outperforms state-of-the-art models like HA-GAN, 3D LDM, and WDM-3D, achieving significantly lower 3D FID, MMD, and Wasserstein distance scores. The model is accepted at ICLR 2026, and the code is publicly available. By enabling direct pixel-level synthesis without quality loss from latent compression, PRDiT could advance medical imaging applications such as data augmentation, denoising, and rare disease simulation, where high-fidelity 3D reconstructions are critical.