Vision Foundry is a self-service platform for image analysis, designed for researchers working with unlabeled or hard-to-label image datasets, especially in biomedicine. It extracts high-dimensional feature representations from unlabeled image data.
At its core is DinoMX, a modular PyTorch training framework for self-supervised representation learning with Vision Transformers (ViTs), building on the DINO and DINOv2 frameworks.
Part of the Kentucky Open Science platform: it runs over the Cresco mesh, with versioned storage for datasets, models, and code in GaiaKeep. See the full stack.
DinoMX replaces convolutional segmentation heads such as U-Net with an attention-map-based segmentation strategy: the model uses its native transformer attention maps to localize and interpret image regions.
The training pipeline is optimized for distributed training (FSDP and DDP), and outputs follow the standard Hugging Face model format for sharing.
The DinoMX framework is available; the self-service Vision Foundry platform is in development. For collaboration, use the information request form.
M. S. Gokmen, M. A. Klusty, E. W. Damron, W. V. Logan, A. D. Mullen, C. N. Leach, E. B. Collier, S. E. Armstrong, V. K. C. Bumgardner. Vision Foundry: A System for Training Foundational Vision AI Models. AMIA Joint Summits on Translational Science Proceedings, 2026, pp. 153-162. arXiv