view article
Figure 1
Schematic representation of the AXIS multi-step transfer learning process. The DINOv2 model trained on 142 000 000 images was used as a starting point (top). LoRA transfer learning on this model was applied using visible light images from the MARCO training dataset to generate the AXIS-foundation model (middle). This model was then independently re-trained with visible and UV images to generate two independent classifiers that operate in visible and UV images, respectively (bottom panel).

IUCrJ
Volume 13| Part 3| May 2026| Pages 235-248
ISSN: 2052-2525
https://doi.org/10.1107/S2052252526001399