Conceptvisual token / 视觉token
Visual Token
A token obtained by discretizing continuous visual features through a vector quantizer, which can participate in language model prediction and generation in the same way as text tokens.
Timeline (last 90 days)
Oct 18
Traditional multimodal models first use a vision encoder (such as ViT) to slice the entire image into a fixed number of patches, map each patch to a visual token, and then feed them together with text tokens into the language model; this process is irreversible.
Unverified50%