Beyond Compression: Quantifying Spectral Accessibility in Vision Representations

Akayou A. Kitessa, Yijun Zhao

Vision-language models map visual features into a shared embedding space through learned projection layers, yet it remains unclear how these transformations alter the structure of…