NV-Generate-MR

A Hounsfield unit means the same thing on every CT scanner, which is why one generative model can plausibly cover the whole body in CT. MR intensities carry no such absolute scale: the same liver looks one way on T1 and another on T2, different again on two vendors' scanners. A generator trained across that mixture without being told which acquisition it is imitating tends to average the contrasts into something that resembles none of them.

NV-Generate-MR makes the acquisition an explicit input instead. Built by NVIDIA with collaborators at the National Institutes of Health, the University of Oxford and the University of Arkansas for Medical Sciences, it is the MR checkpoint of the NV-Generate-CTMR family, released in October 2025. The diffusion model is conditioned on a modality code — unspecified MR, T1, T2 or FLAIR — and on the target voxel spacing, and on nothing else; body region is not an input. Classifier-free guidance is then applied to that modality code during sampling, which is what stops a requested T2 from drifting back toward the corpus mean. The shipped guidance scale of 10, against 0 for the CT checkpoint, is a fair measure of how much pressure the MR case needs.

The checkpoint uses the MAISI-v2 rectified-flow architecture, whose published experiments are CT-only; MAISI-v2 covers that side of the family, and the original DDPM generator is cataloged as MAISI. One capability deliberately does not carry over: no ControlNet branch is released for MR, so this model synthesizes images only, never a paired segmentation mask — the repository ships a training config for one, but no weights. NVIDIA's stated reason is that MR variability is too large to train a single whole-body mask-conditioned model on.

#Key Features

• Contrast chosen at inference: A single integer in the config selects unspecified MR, T1, T2 or FLAIR output, so one set of weights covers the sequences a study actually mixes.
• Four anatomies from one checkpoint: T1 and T2 brain, T2 prostate, T1 breast, and T1 and T2 abdomen, each with a recommended field of view drawn from the median of its training images.
• Thirty-step sampling: Rectified flow transports noise to data along a near-straight path, so an ODE solver reaches a finished volume in 30 evaluations rather than a thousand.
• Geometry as a run parameter: Output volumes of up to 512 x 512 x 128 voxels at voxel spacings from 0.5 to 5.0 mm, written out as NIfTI, DICOM or Nrrd.
• Modest hardware: Inference runs from the downloaded checkpoint on a single 16GB GPU, with no training step and no user data needed.

#Technical Details

The generator is a 240M-parameter 3D UNet with attention blocks, assembled from MONAI components and trained from scratch. It operates in the latent space of a variational autoencoder retrained with eight further CT and MR sources, including abdomen MRI — the MR checkpoint's most substantive departure from its siblings below the diffusion model, and the reason it ships research-only. The diffusion model itself was trained on roughly 16,000 MR volumes from 17 public datasets spanning multiple scanner types (the shared autoencoder saw about 20,000 MRI volumes more broadly), with the per-anatomy composition skewed toward brain T1 (4,659 images) and breast T1 (2,162), and thinner coverage of prostate T2 (898), abdomen T1 (715) and abdomen T2 (78). The runtime is MONAI Core v1.5.0 on PyTorch, tested on A100 and H100 hardware. Because the model is conditioned on spacing rather than anatomy, output quality tracks whether the requested field of view — voxel dimensions times spacing — falls inside the training distribution for that anatomy; numerically valid but out-of-distribution combinations degrade badly, which is why the repository ships a per-anatomy table of recommended values. Contrast-enhanced MRI is outside the model's scope.

#Applications

Synthetic MR volumes are most useful where real ones are hardest to obtain or share: augmenting training sets for rare pathology, standing up segmentation or registration experiments before institutional data access is granted, and circulating realistic examples without patient consent constraints. Prostate and breast MRI are the clearest beneficiaries, being clinically important and poorly represented in public corpora. NVIDIA positions the checkpoint as a base for fine-tuning on a site's own MRI as well as a direct generator.

#Impact

NV-Generate-MR extends a CT-first synthesis framework to a modality where the conditioning problem is harder, shipping a narrower feature set rather than a mask-conditioned pipeline MR data would not support. Two caveats matter for anyone planning to use it. No MR-specific image-quality metric has been published: the FID and 33x acceleration figures in the MAISI-v2 paper were measured on CT, and crediting them to this checkpoint would be a mistake. And the weights ship under the NVIDIA OneWay Non-Commercial License for academic research only, more restrictive than the NVIDIA Open Model License covering the CT and brain-MR checkpoints, even though the surrounding code is Apache-2.0. The release is versioned 0.1 and marked research and development only.