Decoupling Endpoint and Semantic Transition Learning for Zero-Shot Composed Image Retrieval

Researchers propose DeCIR, a new approach to zero-shot composed image retrieval that separates endpoint matching from semantic transition learning to overcome limitations in projection-based methods. The technique uses decoupled text adapters and low-rank directional merging to improve performance on image retrieval tasks without increasing computational complexity at inference time.

DeCIR addresses a fundamental technical challenge in computer vision and machine learning research—the problem of retrieving target images using both a reference image and text modifications. Current projection-based methods struggle with complex semantic changes because they conflate two distinct tasks: matching image endpoints and learning semantic transitions. The innovation lies in recognizing this as a constraint optimization problem rather than a unified learning objective.

The research builds on established techniques in transfer learning and adapter-based fine-tuning, which have become standard in foundation models. By decomposing the problem into separate learning branches that handle endpoint alignment and transition modeling independently, DeCIR reduces the interference between competing objectives. The Low-Rank Directional Merge technique provides an elegant deployment solution, maintaining the efficiency advantages of projection-based approaches while capturing richer semantic information.

This work impacts the broader AI ecosystem by demonstrating how architectural decomposition can improve model performance without sacrificing computational efficiency. For practitioners building image retrieval systems, DeCIR offers a practical alternative to LLM-dependent approaches that require expensive inference resources. The consistent improvements across multiple benchmark datasets (CIRR, CIRCO, FashionIQ, GeneCIS) suggest the approach generalizes well across different domains and image types.

The research direction signals growing sophistication in zero-shot learning, where models must perform complex tasks without task-specific training data. As retrieval systems become increasingly important for e-commerce, content discovery, and multimodal search applications, methods that improve accuracy while reducing computational overhead gain strategic importance.

• →DeCIR decouples endpoint and semantic transition learning to resolve conflicts in projection-based zero-shot composed image retrieval
• →The method uses separate low-rank text adapter branches merged with Low-Rank Directional Merge for efficient deployment
• →Consistent improvements demonstrated across four major benchmark datasets without increasing inference complexity
• →Approach offers lightweight alternative to LLM-based methods while improving performance on complex semantic modifications
• →Research advances the field of zero-shot learning by addressing fundamental architectural limitations in unified learning objectives