Leveraging Generative Representations for Fine-Grained Visual Understanding

Analysis shows that native generative-side representations in multimodal models are more sensitive to fine-grained semantic and spatial changes than understanding-side representations. This finding suggests that generative pathways can improve performance in tasks requiring subtle attribute distinction.