Training-Free Conditional Image Embedding with LVLMs
Published:Dec 26, 2025 04:51
•1 min read
•ArXiv
Analysis
This paper introduces DIOR, a novel, training-free method for generating conditional image embeddings using Large Vision-Language Models (LVLMs). The significance lies in its ability to focus image representations on specific textual conditions without requiring any additional training, making it a versatile and efficient solution. The paper's contribution is particularly noteworthy because it leverages the power of pre-trained LVLMs in a novel way, achieving superior performance compared to existing training-free baselines and even some methods that require training.
Key Takeaways
- •DIOR is a training-free method for generating conditional image embeddings.
- •It leverages Large Vision-Language Models (LVLMs).
- •DIOR outperforms existing training-free baselines.
- •It provides a versatile solution applicable to any image and condition.
Reference
“DIOR outperforms existing training-free baselines, including CLIP.”