Vision encoder
The part of a multimodal model that turns an image into vectors, usually by cutting it into small patches.
Also: image encoder, ViT, CLIP
Most vision encoders are vision transformers, a design Google introduced in 2020 that cuts an image into a grid of small patches and treats each patch like a token. OpenAI's CLIP, released in 2021, trained such an encoder alongside a text encoder on 400 million image and caption pairs so that matching images and texts got similar vectors. CLIP and Google's SigLIP encoders are the starting point for many open vision-language models. Higher image resolution means more patches and more tokens.