MC-LLaVA: Multi-Concept Personalized Vision-Language Model

Overview

• Name: MC-LLaVA
• Type: Multi-Concept Personalized Vision-Language Model
• Paper: MC-LLaVA: Multi-Concept Personalized Vision-Language Model

Model Features

• Multi-Concept Personalization: Through a joint training strategy, MC-LLaVA can integrate multiple concepts in a single training session, achieving multi-concept personalization.
• Utilization of Visual Tag Information: Leverages visual tag information for concept label initialization, enhancing concept representations and accelerating joint training.

Dataset

• Data Source: Images containing multiple characters collected from various movies, with manually generated multi-concept QA samples.
• Dataset Characteristics:
  • Diverse movie genres
  • Diverse question‑answer types

Experimental Results

• Multi-Concept Personalization Response: Through comprehensive qualitative and quantitative experiments, MC-LLaVA demonstrates outstanding multi-concept personalization response capabilities.