CLIP

#clip #vision-language #contrastive #multimodal

Intro

Speeding Up

SigLip

Distillation

• [2310.13355] SILC: Improving Vision Language Pretraining with Self-Distillation
  • GitHub - google-research/silc: [ECCV 2024] Official Release of SILC: Improving vision language pretraining with self-distillation

Improving Representation

• [2410.02746] Contrastive Localized Language-Image Pre-Training

Uses of CLIP

Classification

• ECCV Poster uCAP: An Unsupervised Prompting Method for Vision-Language Models
• [2410.05270] Fine-Tuning CLIP’s Last Visual Projector: A Few-Shot Cornucopia

Data Curation

• DataComp
• DataComp Workshop

Filtering

Deduplication

Noise

Quality Filtering

• Use pretrained CLIP model to filter out pairs with low score

Augmentation