07.10.2026
Our paper Diagnosing Diversity Collapse and Validating Mask-Conditioned Diffusion for Labeled Microtubule Microscopy has been accepted as a spotlight at the Northern Lights Deep Learning Conference (NLDL 2027). The proceedings will appear in PMLR.
The paper presents DiffuMT, a diffusion model that turns a segmentation mask into a realistic microscopy image of microtubules, so every generated image comes with its label. It also shows that standard metrics such as FID miss a failure mode of these models: late in training, they lose diversity. A diagnostic that checks realism and diversity together picks a model whose images domain experts cannot tell apart from real recordings, and a segmentation model trained only on its images performs as well as one trained on real data.
The model and a dataset of real and generated images are available on Hugging Face. Can you tell real from generated? Try it on the project page. The work is a collaboration between HTW Berlin, BHT and the Max Planck Institute for Infection Biology, and follows our SynthMT paper in PLOS Computational Biology.
