Colorist — DenseNet121 on DDI (seed 265017005)

One checkpoint from the model zoo accompanying "Simple, Safe, and Overlooked: Reclaiming Sustainable Domain Generalization with Statistical Color Matching" (MICCAI DEMI 2026, arXiv:2608.18915).

Colorist is a training-free augmentation: it matches each colour channel's first two moments to those of a randomly drawn style image. That is a per-pixel affine map, so it leaves the spatial anatomical layout untouched. The augmentation has no parameters, so there are no weights for it. This repository holds a downstream artefact: a DenseNet121 trained with it on DDI, using RGB mean-standard-deviation matching applied with probability 0.3.

Model Details

Developed by: Sebastian Doerrich, xAILab Bamberg, University of Bamberg
Model type: DenseNet121 image classifier (timm), trained from scratch on DDI
License: MIT. The weights carry the code's license; training used no framework imposing stronger terms. The training data does not — DDI keeps its own license and is not redistributed here.

Sources

Paper: https://arxiv.org/abs/2608.18915
Code: https://github.com/sdoerrich97/colorist
Package: https://pypi.org/project/colorist-aug/
Collection: all 57 checkpoints are grouped in the Colorist collection

Usage

from colorist import load_pretrained_classifier

model, config = load_pretrained_classifier("ddi", seed=265017005)

Needs the optional extra: pip install "colorist-aug[weights]". Preprocess with the normalization in config.json; ImageNet statistics are not what this was trained on.


Dataset	DDI (`ddi`)
Classes	2 (binary-class)
Input	224 x 224 RGB
Seed	`265017005`

Results

Test balanced accuracy on DDI, for every published seed:

Seed	Test balanced accuracy
`71397589`	0.5041
`133560673`	0.4969
`265017005`	0.5000

Mean over seeds 0.5003 (sd 0.0036).

Trained without any augmentation, the same architecture reaches 0.4827 on this dataset.

Read the three seeds together, not this one alone. Every number in the paper is a mean over all three, and on the small covariate-shift benchmarks the spread between them is large. The other two:

Do not average these checkpoints into one. They are trained from scratch with different initialisations, so they occupy different loss basins and averaging their weights cancels their features rather than combining them: measured on BloodMNIST, the three seeds score 0.9844 / 0.9886 / 0.9918 and their weight-average scores 0.1250, which is exactly chance for 8 classes.

Limitations

Photometric shift only. The transform is affine per pixel and changes no geometry, so it does not address geometric or resolution shift.
A research artefact. Not validated for clinical use; not for diagnosis.

Citation

@article{doerrich2026colorist,
  title         = {Simple, Safe, and Overlooked: Reclaiming Sustainable Domain
                   Generalization with Statistical Color Matching},
  author        = {Doerrich, Sebastian and Di Salvo, Francesco and
                   Rai, Shyam Nandan and Lents, Marco and Ledig, Christian},
  year          = {2026},
  eprint        = {2608.18915},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.18915}
}