Tobias Christian Nauen

LUMA reveals that the choice of pretraining objective trumps architectural complexity in driving segmentation quality, challenging existing assumptions in the field.

Image classification accuracy jumps by nearly 3% simply by whispering text embeddings into the ear of a vision model during training.

Stunning visuals from the latest text-to-image models don't guarantee better training data -- in fact, they can make your image classifiers *worse*.