PSI3D brings diffusion AI to 1024³ 3D imaging with slice-wise sampling
PSI3D reconstructs 1024×1024×128 volumes that choked standard diffusion priors.
Diffusion models are powerful image priors for Bayesian inverse problems, but they've been stuck on small, low-dimensional data because training and inference costs explode for large volumes. Now a team led by Wenhan Guo at Johns Hopkins University (with collaborators at Duke and other institutions) proposes PSI3D, a plug-and-play 3D stochastic inference algorithm that finally scales diffusion-based reconstruction to massive 1024×1024×128 volumes. Instead of running a 3D diffusion model, PSI3D uses a Markov chain Monte Carlo (MCMC) approach to reconstruct each 2D slice by sampling from a compact 2D latent diffusion model, drastically cutting memory and compute overhead.
To prevent slice-by-slice artifacts, the algorithm incorporates total variation (TV) regularization stochastically along the concatenation axis. The researchers evaluated PSI3D on optical coherence tomography (OCT) super-resolution, a key medical imaging task. Results show significant improvements in reconstruction quality over both traditional iterative methods and learning-based baselines, while also providing uncertainty quantification through posterior sampling. The method is plug-and-play, meaning it works with existing pretrained 2D diffusion models, making it immediately practical for scientific imaging applications that deal with terabyte-scale volumetric data—without retraining a costly 3D model.
- PSI3D handles 1024×1024×128 volumes (over 134 million voxels) using only 2D latent diffusion models
- Uses MCMC slice-wise sampling with stochastic TV regularization to maintain inter-slice consistency
- Outperforms traditional and learning-based baselines on OCT super-resolution while providing credible posterior samples
Why It Matters
Enables diffusion-prior reconstruction on terabyte-scale scientific/medical volumes where 3D diffusion models are computationally infeasible.