Vision-by-Language for Training-Free Compositional Image Retrieval

Given an image and a target modification (e.g an image of the Eiffel tower and the text "without people and at night-time"), Compositional Image Retrieval (CIR) aims to retrieve the relevant target image in a database. While supervised approaches rely on annotating triplets that is costly (i.e. query image, textual modification, and target image), recent research sidesteps this need by using large-scale vision-language models (VLMs), performing Zero-Shot CIR (ZS-CIR). However, state-of-the-art approaches in ZS-CIR still require training task-specific, customized models over large amounts of image-text pairs. In this work, we propose to tackle CIR in a training-free manner via our Compositional Image Retrieval through Vision-by-Language (CIReVL), a simple, yet human-understandable and scalable pipeline that effectively recombines large-scale VLMs with large language models (LLMs). By captioning the reference image using a pre-trained generative VLM and asking a LLM to recompose the caption based on the textual target modification for subsequent retrieval via e.g. CLIP, we achieve modular language reasoning. In four ZS-CIR benchmarks, we find competitive, in-part state-of-the-art performance -improving over supervised methods. Moreover, the modularity of CIReVL offers simple scalability without re-training, allowing us to both investigate scaling laws and bottlenecks for ZS-CIR while easily scaling up to in parts more than double of previously reported results. Finally, we show that CIReVL makes CIR human-understandable by composing image and text in a modular fashion in the language domain, thereby making it intervenable, allowing to post-hoc re-align failure cases. Code available at github.com/ExplainableML/Vision by Language. INTRODUCTION Compositional Image Retrieval (CIR) necessitates a nuanced coupling between the image content and the semantics of the textual query to retrieve a new image that accurately embodies the relevant image elements and the modifications described in the textual query. To achieve this, previous works require curated triplets (query image, modifying text, target image) to train a specific CIR system. However, annotating such triplets is both difficult and labor-intensive. To tackle this problem, recent research proposed Zero-Shot CIR (ZS-CIR) (Saito et al., 2023; Baldrati et al., 2023) . Based on largescale pre-trained vision-language models (VLMs) (e.g. CLIP (Radford et al., 2021)), these methods use image-caption pairs to train textual inversions (Gal et al., 2023; Cohen et al., 2022) mapping images to text tokens. A static template merges tokens and textual modifications to obtain target captions, performing CIR without explicit supervision. Thus, even when leveraging large-scale VLMs, ZS-CIR methods still train additional mapping networks on large image-caption datasets. In this work, we propose to achieve training-free ZS-CIR by leveraging ubiquitously available, offthe-shelf models already trained with large-scale training data. Our Compositional Image Retrieval through Vision-by-Language (CIReVL) follows the vision-by-language paradigm (