DiffuMT

Mask-conditioned diffusion dataset of in-vitro IRM microtubule microscopy. Each labelled item is a triplet — a pixel-precise binary mask, the real recording, and a mask-conditioned synthetic counterpart — with images that domain experts cannot reliably distinguish from real ones in a 2AFC study.

Splits: test (200) is live now. train (2961) and unlabeled (1336, real only) are planned additions, not yet released. Filenames are anonymised to diffumt_{counter}; mask/real/synthetic of one item share the same id.

2AFC challenge (integrated into the project page): https://huggingface.co/spaces/HTW-KI-Werkstatt/DiffuMT

Citation

Accepted at NLDL 2027 (Spotlight), to appear in PMLR. Preprint on arXiv.

@inproceedings{koddenbrock2027diffumt,
  title     = {Diagnosing Diversity Collapse and Validating Mask-Conditioned
               Diffusion for Labeled Microtubule Microscopy},
  author    = {Koddenbrock, Mario and Rapp, Frederic and Reber, Simone and
               Rodner, Erik},
  booktitle = {Proceedings of Machine Learning Research (PMLR)},
  year      = {2027},
  note      = {Northern Lights Deep Learning Conference (NLDL), Spotlight},
  eprint    = {2610.09957},
  archivePrefix = {arXiv}
}