CLIP-IN: Enhancing Fine-Grained Visual Understanding in CLIP via Instruction Editing Data and Long Captions

2025-08-04

AI Summary

Despite their success, vision-language models like CLIP exhibit significant limitations in fine-grained visual understanding. To address this, we propose a CLIP enhancement framework specifically designed for fine-grained semantic discrimination. First, we construct high-quality hard negative image-text pairs using image manipulation–oriented instruction data and introduce a symmetric contrastive loss to strengthen discrimination of subtle visual distinctions. Second, we incorporate a long-text description modeling module equipped with rotary position encoding to better capture complex visual-semantic contextual dependencies. Evaluated on fine-grained benchmarks such as MMVP, our method achieves substantial performance gains while preserving zero-shot classification and cross-modal retrieval capabilities—and notably mitigates visual hallucination in multimodal large language models. The core innovations lie in (i) an instruction-driven hard negative mining mechanism and (ii) a long-context-aware text encoding strategy.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Despite the success of Vision-Language Models (VLMs) like CLIP in aligning vision and language, their proficiency in detailed, fine-grained visual comprehension remains a key challenge. We present CLIP-IN, a novel framework that bolsters CLIP's fine-grained perception through two core innovations. Firstly, we leverage instruction-editing datasets, originally designed for image manipulation, as a unique source of hard negative image-text pairs. Coupled with a symmetric hard negative contrastive loss, this enables the model to effectively distinguish subtle visual-semantic differences. Secondly, CLIP-IN incorporates long descriptive captions, utilizing rotary positional encodings to capture rich semantic context often missed by standard CLIP. Our experiments demonstrate that CLIP-IN achieves substantial gains on the MMVP benchmark and various fine-grained visual recognition tasks, without compromising robust zero-shot performance on broader classification and retrieval tasks. Critically, integrating CLIP-IN's visual representations into Multimodal Large Language Models significantly reduces visual hallucinations and enhances reasoning abilities. This work underscores the considerable potential of synergizing targeted, instruction-based contrastive learning with comprehensive descriptive information to elevate the fine-grained understanding of VLMs.

Problem

• Enhancing fine-grained visual understanding in CLIP models
• Distinguishing subtle visual-semantic differences effectively
• Reducing visual hallucinations in multimodal language models

Innovation

• Instruction-editing datasets for hard negative pairs
• Symmetric hard negative contrastive loss training
• Long captions with rotary positional encodings