Composed Image Retrieval (CIR) is an image retrieval task where users provide a reference image along with modification text to retrieve a target image. Zero-shot CIR (ZS-CIR) attracts significant research interest owing to its strong generalization capability and independence from labeled training data. Most ZS-CIR methods employ late fusion and textual inversion, but these approaches fail to precisely convert the reference image and modification text into a target-aligned query. Consequently, the final query may retain redundant information—such as elements present in the reference image but absent in the target image. To address these limitations, a training-free ZS-CIR method called Test-time Filtering (TTF) is proposed in this paper, leveraging Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs). Specifically, an MLLM generates captions for reference images, which, along with the modification text, are fed into an LLM to produce target image descriptions. This converts the final query into plain text. Using these results, an MLLM filters the selected candidate images to select positive and negative samples. Finally, the similarity scores of candidate images matching positive samples are amplified, while those of negative samples are suppressed. The proposed method is evaluated on three ZS-CIR datasets—CIRR, FashionIQ, and CIRCO—with experimental results demonstrating superior performance over prior ZS-CIR approaches. The source code is available at https://github.com/After-lifes/TTF.

Compared to supervised composed image retrieval (CIR), which requires a large number of manually labeled triplets (reference image, modification text, target image) for training, Zero-shot Composed Image Retrieval (ZS-CIR) only needs easily available image-caption pairs for training. Prior works for ZS-CIR have primarily used textual inversion to convert images from image-caption pairs into pseudo-words using a pre-trained Vision-language Model (VLM). These pseudo-words are then concatenated to a template, such as “a photo of $”. We find that this fixed template representation hinders the generalization ability of ZS-CIR, and that converting an image into a single pseudo-token fails to accurately represent the information contained in the image. Additionally, given the high performance demonstrated by traditional supervised CIR after triplet training, we aim to achieve a similar structure in ZS-CIR. In this paper, we propose a novel ZS-CIR paradigm based on pseudo-triplets and flexible prompts, rather than fixed templates. Specifically, we fine-tune the vision encoder and split the caption to generate pseudo-triplets using random patch masking. The obtained masked image tokens and corresponding split captions are fed into a Q-former, which is guided by the learned queries and generates queries containing both image and caption information. Finally, the learned queries are concatenated with the split caption and fed into the text encoder to obtain a composed query. Extensive experiments on commonly used ZS-CIR benchmarks such as CIRR, FashionIQ, and CIRCO demonstrate that our method outperforms previous ZS-CIR approaches. The source code is available at https://github.com/After-lifes/PFP.

With the explosive growth of multi-modal data such as video, images, and text on the Internet, cross-modal retrieval has received extensive attention, especially the deep hashing method. Compared with the real-value method, deep hashing has shown promising prospects due to its low memory consumption and high searching efficiency. However, most existing studies have difficulties in effectively utilizing the raw image-text pairs to generate discriminative feature representations. Moreover, these methods ignore the latent relationship between different modalities and fail to construct a robust similarity matrix, resulting in suboptimal retrieval performance. In this paper, we focus on the unsupervised cross-modal hashing tasks and propose a Self Attentive CLIP Hashing (SACH) model. Specifically, we construct the feature extraction network by employing the pre-trained CLIP model, which has shown excellent performance in zero-shot tasks. Besides, to fully exploit the semantic relationships, an attention module is introduced to reduce the disturbance of redundant information and focus on important information. On this basis, we construct a semantic fusion similarity matrix that capable of preserving the original semantic relationships from different modalities. Extensive experiments show the superiority of SACH compared with recent state-of-the-art unsupervised hashing methods.