Skip to content

ml-ferret

Ground referring expressions in multimodal inputs

Visit official site
Category AI models

What ml-ferret is

The model is described as an end-to-end multimodal language model that accepts many forms of reference and performs precise grounding.

What you can do with it

  • Accept references in several forms
  • Ground regions in a multimodal setting
  • Handle fine-grained and open vocabulary terms
  • Evaluate results with Ferret-Bench

Model notes

  • A hybrid region representation is used
  • A spatial aware visual sampler is used
  • The GRIT dataset holds about 1.1 million samples

Pros & cons

✓ What we like

  • The GRIT dataset contains about 1.1 million samples
  • Ferret-Bench provides a multimodal evaluation benchmark

! What to watch out for

  • The described model is end-to-end by design
  • Grounding is tied to the given referring expressions

FAQ

What is included besides the model?

The GRIT dataset of about 1.1 million samples and the Ferret-Bench benchmark.

Which components are named?

Hybrid region representation and a spatial aware visual sampler.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review