Related Projects

MR simulators are not one kind of thing, and the first question about any of them is what it is for. Two families answer different questions, and a package is usually only comparable to the others in its own.

Image formation. Given a pulse sequence with its real waveforms and a phantom with positions, what raw data comes off the scanner? These integrate the Bloch equation over isochromats that have coordinates, so gradients, off-resonance maps, coil sensitivities, motion and the readout itself are all in the model. JEMRIS [1], MRiLab [2], KomaMRI.jl [3], CMRsim [4] and MRzero-Core [5] answer this question.

Signal evolution. Given a train of pulses and a tissue, what does one voxel record? No coordinates, no k-space, no image: the sequence is a stream of events and the answer is a curve per tissue. That is what a dictionary, a fit, a Cramer-Rao bound and a model-based reconstruction are built from, and it is what an extended phase graph computes in a fraction of the states an isochromat ensemble needs. sycomore [6], epgpy [7], EpyG [8], mri-sim-py [9], snapMRF [10] and TorchSim answer this one. BlochSimulators.jl [11] answers both, from the same sequence description.

TorchSim is the second kind. It has no gradient waveforms, no phantom coordinates and no encoding operator of its own – How TorchSim runs it lists what that rules out. What it adds instead is the derivative: the same kernels that produce a signal produce its Jacobian with respect to tissue, and its gradient with respect to the sequence, which is what the estimators, the model-based reconstruction and the sequence design on these pages consume.

The landscape

Read as of August 2026, from each project’s own documentation. A blank is not a criticism: a scanner simulator has no reason to carry a dictionary Jacobian, and a signal-model simulator has no reason to carry a coil.

Project	Model	Implementation	Hardware	Derivatives	Sequence input
JEMRIS	Isochromat, ODE solver (CVODE)	C++, MPI	CPU clusters; a GPU port published in 2025 [12]	–	XML, GUI
MRiLab	Discrete spin, multi-pool exchange tissue model	MATLAB front end, C++/CUDA kernels	CPU threads, CUDA	–	GUI
KomaMRI.jl	Isochromat, operator splitting and Magnus expansions	Julia	CPU threads, CUDA, AMDGPU, Metal, oneAPI	–	Pulseq .seq, Julia, GUI
CMRsim	Isochromat, and analytic signal models	Python, TensorFlow 2	CPU, GPU	–	Python
MRzero-Core	Phase distribution graphs, and Bloch	Python, PyTorch, Rust core	CPU, CUDA	Automatic, reverse mode	Pulseq .seq, PyPulseq
BlochSimulators.jl	Isochromat and EPG	Julia	CPU threads, distributed, CUDA	Finite differences, in MR-STAT [11]	Julia
sycomore	Bloch, and EPG: regular, discrete, discrete 3D	C++ core, Python bindings	One CPU core	–	Python
epgpy	EPG, with 3D gradients and multi-compartment exchange	Python, NumPy or CuPy	CPU, CUDA through CuPy	Analytic, first and second order	Python
EpyG	EPG	Python	CPU	–	Python
mri-sim-py	EPG	Python, PyTorch	CPU, CUDA	Automatic, reverse mode	Python
snapMRF	EPG, with matching	CUDA C	CUDA	–	Command line
TorchSim	EPG, and closed forms	Python, PyTorch, C++ and Triton kernels	CPU threads, CUDA, several cards	Automatic: forward, reverse, and forward over reverse	Python, or a description

Where each one is the better tool

A sequence you are developing, and the image it makes. KomaMRI, JEMRIS, CMRsim or MRzero-Core. They read the waveforms you will play, carry the spins through them at their coordinates, and hand back k-space. TorchSim reads a sequence as events rather than waveforms, and stops at the signal.

One curve, read interactively, with the model in front of you. sycomore or epgpy. Both are a few milliseconds for one tissue with nothing to warm up, and sycomore’s units make an expression read like the paper it came from. TorchSim resolves the structure of a sequence before it runs one, which is seconds it does not repay until there is a dictionary to sweep or a loop to run.

A dictionary, a fit, a design loop, or a map solved from k-space. TorchSim. The batching across tissues, the derivative, and the estimators and reconstruction that consume it are the point of the package; benchmarks/ in the repository measures the first two against the alternatives above and states the agreement between them.

A steady state that has a closed form. Its expression, which is faster than any of these – and several ship here as closed-form simulators.

References

[1]

Stöcker, T., Vahedipour, K., Pflugfelder, D., Shah, N. J., “High- performance computing MRI simulations”, Magnetic Resonance in Medicine (2010). https://doi.org/10.1002/mrm.22406

[2]

Liu, F., Velikina, J. V., Block, W. F., Kijowski, R., Samsonov, A. A., “Fast realistic MRI simulations based on generalized multi-pool exchange tissue model”, IEEE Transactions on Medical Imaging (2017). https://doi.org/10.1109/TMI.2016.2620961

[3]

Castillo-Passi, C., Coronado, R., Varela-Mattatall, G., Alberola-López, C., Botnar, R., Irarrazaval, P., “KomaMRI.jl: An open-source framework for general MRI simulations with GPU acceleration”, Magnetic Resonance in Medicine (2023). https://doi.org/10.1002/mrm.29635

[4]

Weine, J., McGrath, C., Dirix, P., Stoeck, C. T., Kozerke, S., “CMRsim – A python package for cardiovascular MR simulations incorporating complex motion and flow”, Magnetic Resonance in Medicine (2024). https://doi.org/10.1002/mrm.30010

[5]

Endres, J., Weinmüller, S., Dang, H. N., Zaiss, M., “Phase distribution graphs for fast, differentiable, and spatially encoded Bloch simulations of arbitrary MRI sequences”, Magnetic Resonance in Medicine 92.3 (2024), pp. 1189-1204. https://doi.org/10.1002/mrm.30055

[6]

Lamy, J., “sycomore: an MRI simulation toolkit”. lamyj/sycomore

[7]

Baudin, P., “epgpy: EPG simulations in Python”. py-baudin/epgpy

[8]

Brenner, D., “EpyG: extended phase graphs in Python”. brennerd11/EpyG

[9]

“mri-sim-py.epg: a GPU-accelerated extended phase graph algorithm for differentiable optimization and learning”. utcsilab/mri-sim-py.epg

[10]

Wang, D., Ostenson, J., Smith, D. S., “snapMRF: GPU-accelerated magnetic resonance fingerprinting dictionary generation and matching using extended phase graphs”, Magnetic Resonance Imaging 66 (2020), pp. 248-256. https://doi.org/10.1016/j.mri.2019.11.015

[11] (1,2)

van der Heide, O., Sbrizzi, A., Bruijnen, T., van den Berg, C. A. T., “GPU-accelerated Bloch simulations and MR-STAT reconstructions using the Julia programming language”, Magnetic Resonance in Medicine (2024). https://doi.org/10.1002/mrm.30074

[12]

“GPU-accelerated JEMRIS for extensive MRI simulations”, Magnetic Resonance Materials in Physics, Biology and Medicine (2025). https://doi.org/10.1007/s10334-025-01281-z