LiFT: 3D Medical Image Generation via Inter-slice Feature Trajectories

Researchers propose LiFT (Lifted Inter-slice Feature Trajectories), a framework for high-resolution 3D medical image generation that avoids computationally expensive volumetric models. Instead of modeling the full 3D distribution, LiFT factorizes synthesis into per-slice 2D generation and inter-slice trajectory learning, treating a volume as an ordered feature trajectory capturing anatomical changes across depth. A tri-planar drifting loss aligns generated slice trajectories with real volumes for unconditional generation; for paired translation, a bidirectional z-context mixer trained against registered targets provides through-plane coherence. The method addresses the challenge of preserving anatomical consistency in the third dimension when using efficient 2D slice generators. The paper is available on arXiv under ID 2605.19060.