zero-shot learning

In this article, we present Long-CLIP, a fine-tuning method for CLIP that maintains original capabilities through two new strategies: (1) preserving knowledge via positional embedding stretching and (2) matching CLIP features' primary components efficiently.