TIGER-AI-Lab/VLM2Vec

Extends unified multimodal embeddings to videos and visual documents via instruction-guided contrastive training on Qwen2-VL backbones, enabling cross-modal retrieval and classification across diverse visual formats. The framework introduces MMEB-V2, a 78-task benchmark spanning image, video, and document modalities for systematic evaluation. Integrates with Hugging Face for model checkpoints and datasets, and has been integrated into vLLM for production inference.

592 stars.