Computer scientists have developed a lightweight end-to-end transformer framework named SeGCIR to improve zero-shot composed image retrieval, which retrieves target images using multimodal queries composed of reference images and modification text without requiring annotated triplets from target benchmarks. Existing zero-shot composed image retrieval methods frequently rely on multimodal fusion, pseudo-token projection, synthetic supervision, or query reformulation driven by large language models and multimodal large language models. Although those established techniques are effective, the researchers note that they can introduce additional training or inference complexity while depending on prompt design, external generative reasoning, or large-scale supervision.
To address these operational limitations, the research team designed SeGCIR to learn target-oriented composed query representations using only 32,000 training triplets. The framework carries out retrieval tasks directly without relying on inference reasoning from large language models or multimodal large language models. The architecture incorporates two core modules designated as SGM and ProNet, which work together to refine edit-aware global semantics, fine-grained visual-language cues, and reference-text signals into a stronger composed representation.
The system additionally introduces a training-time calibrator named QBM alongside a candidate-aware ranking strategy applied during final retrieval operations. In experimental evaluations conducted across the CIRR, FashionIQ, and CIRCO benchmarks, the proposed framework demonstrated strong and consistent retrieval performance across test settings. The researchers concluded that well-grounded composed representations can generalize effectively without depending on large-scale synthetic supervision or generative multimodal reasoning during inference.
Subject of Research: Zero-shot composed image retrieval using lightweight transformer architectures
Article Title: Semantic grounding-driven zero-shot composed image retrieval
Article References: Rachabathuni, P. K., Ciamarra, A., Caldelli, R., & Bertini, M. (2026). Semantic grounding-driven zero-shot composed image retrieval. Multimedia Tools and Applications, 85(11), Article 852. https://doi.org/10.1007/s11042-026-22000-3
Image Credits: AI Generated
DOI: 10.1007/s11042-026-22000-3
Keywords: Zero-Shot Composed Image Retrieval, SeGCIR, Multimodal Query, Transformer Framework, SGM Module, ProNet Module, QBM Calibrator, Candidate-Aware Ranking, CIRR Benchmark, FashionIQ Benchmark, CIRCO Benchmark, Image Retrieval, Visual-Language Grounding, Multimedia Indexing, Computer Vision, Pavan Kartheek Rachabathuni
News Source: Denise Maddox. (November 12, 2026). Researchers Introduce SeGCIR Framework to Improve Zero-Shot Composed Image Retrieval. Scienmag.



