Cross-domain validation of late-training diversity collapse

Demonstrate whether the late-training diversity-collapse failure mode identified for mask-conditioned diffusion on microtubule microscopy also occurs in other biomedical imaging domains, including computed tomography and magnetic resonance imaging.

Background

The paper identifies a late-training failure mode in which realism metrics continue to improve while the diversity of generated appearances collapses. The evidence is obtained in a single biomedical domain: mask-conditioned diffusion for labeled microtubule microscopy.

The authors suggest that the mechanism may extend to other spatially conditioned biomedical generators, but do not establish this empirically for computed tomography or magnetic resonance imaging. Cross-domain experiments are therefore needed to determine whether the diagnostic and the observed collapse generalize beyond microtubule microscopy.

References

Finally, while the mechanism likely extends to other spatially conditioned generators, demonstrating the same failure mode across additional biomedical domains such as \gls{ct} and \gls{mri} remains future work.

— Diagnosing Diversity Collapse and Validating Mask-Conditioned Diffusion for Labeled Microtubule Microscopy  (2610.09957 - Koddenbrock et al., 7 Oct 2026) in Section 5, Discussion, paragraph “Limitations”