[ACL'26] TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval

Zixu Li1, Yupeng Hu1*, Zhiheng Fu1, Zhiwei Chen1, Yongqi Li2, Liqiang Nie3
1School of Software, Shandong University,
2Department of Computing, Hong Kong Polytechnic University,
3School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen)
*Corresponding author.
Paper Coming Soon Complete source code

Abstract

MY ALT TEXT

Traditional CIR and Baseline Performance in Multi-Modification Scenarios

MY ALT TEXT

(a) Example of traditional CIR, and (b) Performance comparison of representative baselines on CIR datasets in original and multi-modification scenarios (all models are trained on original FashionIQ).


Data Generation Pipeline

MY ALT TEXT

Pipeline of the construction of our proposed multi-modification CIR datasets.


Framework: Text-oriented Entity Mapping Architecture (TEMA)

MY ALT TEXT

Overall architecture of our proposed TEMA.

Experiment

MY ALT TEXT

Performance comparison on M-FashionIQ and M-CIRR relative to R@K(%). The overall best results are in bold, while best results over baselines are underlined. The Avg metric in M-CIRR denotes (R@5 + Rsubset@1) / 2.


MY ALT TEXT

The qualitative results for the PA module, which shows the MMT and corresponding summary. The to-be-modified entities are colored.


MY ALT TEXT

Attention visualization results for the reference image on M-FashionIQ by the PA-generated summary.


MY ALT TEXT

Attention visualization results for the reference image on M-CIRR by the PA-generated summary.


BibTeX


          @inproceedings{TEMA,
            title={TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval},
            author={Li, Zixu and Hu, Yupeng and Fu, Zhiheng and Chen, Zhiwei and Li, Yongqi and Nie, Liqiang},
            booktitle={Proceedings of the Association for Computational Linguistics (ACL)},
            year={2026}
          }