Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion

TL;DR

This paper reveals intra-modal misalignment in CLIP caused by its inter-modal contrastive loss and proposes modality inversion techniques to improve intra-modal task performance, highlighting the importance of intra-modal constraints.

Contribution

It introduces modality inversion methods to expose intra-modal misalignment in CLIP and demonstrates how addressing this improves intra-modal task performance.

Findings

Intra-modal retrieval performance improves with inter-modal approaches.

Intra-modal tasks are negatively impacted when approached inter-modally.

Incorporating intra-modal constraints reduces intra-modal misalignment.

Abstract

Pre-trained multi-modal Vision-Language Models like CLIP are widely used off-the-shelf for a variety of applications. In this paper, we show that the common practice of individually exploiting the text or image encoders of these powerful multi-modal models is highly suboptimal for intra-modal tasks like image-to-image retrieval. We argue that this is inherently due to the CLIP-style inter-modal contrastive loss that does not enforce any intra-modal constraints, leading to what we call intra-modal misalignment. To demonstrate this, we leverage two optimization-based modality inversion techniques that map representations from their input modality to the complementary one without any need for auxiliary data or additional trained adapters. We empirically show that, in the intra-modal tasks of image-to-image and text-to-text retrieval, approaching these tasks inter-modally significantly improves performance with respect to intra-modal baselines on more than fifteen datasets. Additionally, we demonstrate that approaching a native inter-modal task (e.g. zero-shot image classification) intra-modally decreases performance, further validating our findings. Finally, we show that incorporating an intra-modal term in the pre-training objective or narrowing the modality gap between the text and image feature embedding spaces helps reduce the intra-modal misalignment. The code is publicly available at: https://github.com/miccunifi/Cross-the-Gap.