🌊 DRIFT: Difficulty-aware Rectified Flows for Through-plane MRI Super-Resolution

Super-resolve thick-slice MRI scans using a two-stage rectified flow framework with physics-aware adaptive inference.

Paper: DRIFT (ECCV 2026) | Code: GitHub | Model: yoonseokchoi/drift-hcp

1 6
Examples

How it works

  1. Stage 1 (APN): A 2D U-Net maps the low-resolution input to a coarse high-resolution estimate using slice-thickness conditioning (Ï„ = 1/T).
  2. Stage 2 (RF): A rectified flow model refines the coarse prediction via ODE integration, with adaptive step count based on PAD (Physics-Aware Difficulty = 1 − T_hr/T_lr).
  3. Sliding window inference with 128×128 patches and Gaussian blending handles arbitrary image sizes.

Upload any brain MRI slice (axial, coronal, or sagittal). The demo will simulate thick-slice degradation and reconstruct the high-resolution version.