A Simple yet Efficient Multimodal Feature Modulation Adapter for Visual Grounding

Visual Grounding (VG) aims to establish fine-grained alignment between visual content and natural language, playing a crucial role in multimodal reasoning systems. Existing approaches often rely on full fine-tuning of large-scale pretrained models to achieve visual-text alignment, which incurs substantial computational costs as model size grows. To address this challenge, we propose a lightweight Multimodal Feature Modulation (MFM) adapter for visual grounding that enhances visual-textual alignment while significantly reducing trainable parameters. MFM consists of two cascaded modules: Self-Aware Text Modulation (SATM) and Text-Aware Visual Modulation (TAVM). Both modules leverage a shared gating mechanism to refine features outside the backbone, thus preserving the pretrained representations. Our method updates 3.22M parameters (2.24% of the backbone), yet surpasses full-parameter fine-tuning with a 4.28% average accuracy improvement on benchmark datasets.