What does CLIP stand for and what is its key contribution?

**CLIP** stands for Contrastive Language-Image Pretraining. It is the model from OpenAI that taught the field how to put images and text into the same vector space. The recipe has three parts. First, you collect a huge dataset of (image, caption) pairs from the web. The original CLIP used about 400M of them. Second, you train two encoders side by side: a vision encoder that turns an image into a vector, and a text encoder that turns a caption into a vector. Third, you train them with a contrastive loss that says matched pairs should have high cosine similarity and mismatched pairs should have low cosine similarity. After training, the two encoders run independently. You can embed only text, only images, or compare across modalities. The key contribution is the shared space itself. Before CLIP, you needed labeled image datasets and a separate classifier for every task. After CLIP, you could embed any caption string and use it as a zero-shot classifier, or do text to image search out of the box.

**CLIP** is Contrastive Language-Image Pretraining, OpenAI's 2021 model that introduced the modern recipe for joint vision-language embeddings. The architecture is two encoders trained in parallel: a vision encoder (originally a ViT or ResNet variant) and a text encoder (a transformer). The training data is roughly 400M (image, caption) pairs scraped from the web. The objective is a symmetric InfoNCE loss. For each batch, maximize cosine similarity between matched (image, caption) pairs and minimize cosine against every other text in the batch (and symmetrically for text to image). The key contribution is the SHARED embedding space. Image embeddings and text embeddings live in the same coordinate system, so you can compare them directly with cosine similarity. That unlocks three things that were hard before. Text to image search (encode a query string, do nearest-neighbor over image vectors), zero-shot image classification (encode label strings, pick the highest-cosine label). Multimodal RAG (retrieve images by text or vice versa). The CLIP recipe spawned a family: SigLIP swapped InfoNCE for a sigmoid loss that scales better with batch size, OpenCLIP retrained on open data, and EVA-CLIP pushed scale further. The pattern of two encoders into a shared space is now the default for any cross-modal retrieval system in 2026.