metadata
license: mit
LongD-CLIP
Retaining Knowledge and Enhancing Long-Text Representations in CLIP through Dual-Teacher Distillation
📄 CVPR 2025
This repository provides resources for our CVPR 2025 paper:
"Retaining Knowledge and Enhancing Long-Text Representations in CLIP through Dual-Teacher Distillation"
🔍 Introduction
Our work focuses on improving CLIP’s ability to handle long-text inputs while retaining its original knowledge.
We propose a Dual-Teacher Distillation framework that:
- Retains knowledge from the original CLIP,
- Enhances long-text representations through teacher guidance,
This work extends the research line of Long-CLIP and further advances long-text representation learning in multimodal models.
👉 The implementation can also refer to LongD-CLIP.
🚀 Resources
- Paper: CVPR 2025 proceedings
- Model Weights: Hugging Face – LongD-CLIP
- Related Codebase: Long-CLIP