While simulation plays a crucial role in high energy physics, it also consumes a significant fraction
of the available computational resources, with these computing pressures being set to increase
drastically for the upcoming high luminosity phase of the LHC and for future colliders. At the same
time, the significantly higher granularity present in future detectors increases the physical accuracy
required of a surrogate simulator. Machine learning methods based on deep generative models hold
promise to provide a computationally efficient solution, while retaining a high degree of physical
fidelity. Significant strides have already been taken towards developing these models for the
generation of particle showers in highly granular calorimeters, the subdetector which constitutes
the most computationally intensive part of a detector simulation. However, to apply these models
to a general detector simulation, methods must be developed to cope with particles incident at
various points and under varying angles in the detector. This contribution will address steps taken
to tackle the challenges faced when applying these simulators in more general scenarios, as well as
the effects on physics observables after interfacing with reconstruction algorithms. In particular,
results achieved with bounded information bottleneck and normalising flow architectures based
on regular grid geometries, as well as a more flexible diffusion model using point clouds, will be
discussed. Combined with progress on integrating these surrogate simulators into existing full
simulation chains, these developments bring an application to benchmark physics analyses closer.