Can Language Unlock Better Compression?

The Human Intuition

  • Humans naturally prioritize high-level semantics, describing the "what" before the "how many pixels".
  • Text captions provide a concise, meaningful summary, distilling the essence of an image.
  • Descriptive text is now abundant and easy to obtain for most images via modern vision-language models.

A New Paradigm

Leveraging language to sidestep the traditional trade-off between compression and quality.

Dense Pixel Data
Compact Language Tokens
Higher Quality at Lower Bitrates