Vision-language pretraining has driven progress in image-text retrieval and zero-shot recognition by learning transferable cross-modal representations. However, the authors of a new arXiv paper argue that sentence-level alignment alone is insufficient for compositional understanding, where relationships between objects, attributes, and actions matter.
To address this, they introduce Flow-CLIP, which performs image-conditioned residual scoring for compositional image-text matching. Rather than relying only on a global sentence embedding, the method conditions the scoring process on the image itself, allowing for finer-grained comparisons.
The paper is a new submission on arXiv, but the abstract is cut off before details of experiments or results are given. Still, the proposed direction points to a growing focus on compositional reasoning in multimodal models.