Out-of-Distribution Adaptation

NCUT performs well on out-of-distribution (OOD) datasets, demonstrating robust generalization capabilities across diverse image distributions.

Training-Free Runtime Adaptation

NCUT adapts to new data at inference time without requiring retraining or fine-tuning—the backbone model remains frozen. (The backbone model is DINO in the examples below.)

The graph affinity matrix automatically adapts to input images, even when they are out-of-distribution. As nodes in the input set contrast against each other, the resulting affinity matrix (analogous to a kernel matrix) dynamically adjusts to accommodate OOD images.

---

Example 1

Example 2

Example 3

Example 4

Example 5