VolDiT

Controllable Volumetric Medical Image Synthesis
with Diffusion Transformers

MICCAI 2026

Marvin Seyfarth Salman Ul Hassan Dar Yannik Frisch Philipp WildNorbert Frey Florian André Sandy Engelhardt

University Hospital Heidelberg Institute for Artificial Intelligence in Cardiovascular Medicine

Paper Code Cite

High-resolution medical synthesis

Synthetic CT examples from LUNA16 and TaviCT.

Explore the full volume

Rotate, cut through, and inspect synthetic anatomy in three dimensions.

Loads on request · about 13 MB for the first volume and renderer.

Browser preview preserves the full field of view. Powered by NiiVue.

How VolDiT works

Anatomy, modeled in three dimensions

Native 3D tokens. Global context. Spatial control.

VolDiT framework: a volumetric autoencoder compresses a CT scan, a 3D diffusion transformer denoises latent patches, and a timestep-gated control adapter adds segmentation-mask guidance. — VolDiT framework · Figure 1View full figure

Compress the volume
A volumetric autoencoder maps a CT scan into a compact 3D latent space.
Learn global structure
A diffusion transformer models spatial relationships across native 3D tokens.
Guide the anatomy
A timestep-gated adapter adds segmentation guidance to a frozen backbone.

VolDiT combines compact volumetric representation with global self-attention. The same backbone supports unconditional generation and synthesis guided by segmentation masks.

See the evidence

Generation quality, measured

Results from the MICCAI 2026 paper · 100 generated volumes per comparison.

**LUNA16**3D MedicalNet features
Model	FID ↓	Precision ↑	Recall ↑
U-Net LDM	0.031	0.44	0.94
HA-GAN	0.126	1.00	0.00
VolDiT (L, p=4)	0.004	0.91	0.90

**TaviCT**2.5D ImageNet features
Model	FID ↓	Precision ↑	Recall ↑
U-Net LDM	36.8	0.95	0.63
HA-GAN	155.5	1.00	0.00
VolDiT (L, p=2)	21.4	0.94	0.73

FID uses different feature extractors; compare values only within each dataset.

Spatial control from segmentation masks

Median Dice on TaviCT · Learned timestep gate

Heart

VolDiT: 0.94
U-Net: 0.89

Aorta

VolDiT: 0.92
U-Net: 0.83

Fixed-size volumes and dataset-specific autoencoders. Synthetic examples illustrate generation, not clinical validity.

Full results and ablations

Resources & next steps

Build on VolDiT

Paper, implementation, and directions for further research.

Research resources

Cite this work

@inproceedings{seyfarth2026volditcontrollablevolumetricmedical,
  title = {VolDiT: Controllable Volumetric Medical Image Synthesis with
           Diffusion Transformers},
  author = {Seyfarth, Marvin and Dar, Salman Ul Hassan and Frisch, Yannik
            and Wild, Philipp and Frey, Norbert and Andr{\'e}, Florian and
            Engelhardt, Sandy},
  booktitle = {Medical Image Computing and Computer Assisted Intervention --
               MICCAI 2026},
  year = {2027},
  publisher = {Springer Nature Switzerland},
  series = {Lecture Notes in Computer Science},
  volume = {16890},
  pages = {596--606},
  doi = {10.1007/978-3-032-38189-7_57}
}

Ongoing research VolDiT v2

Flow matching

Alternative objectives for training and sampling.

Spacing-aware positions

Position encodings that reflect volume geometry.

Latent normalization

Channel-wise scaling and controlled ablations.

These directions are under investigation; results above refer to the MICCAI 2026 paper.