Skip to content

Migician

Locate targets across multiple images with natural language

Visit official site

What Migician is

The model is described as a multimodal large language model for multi-image grounding from a natural language processing lab at Tsinghua University. The material states a new training framework and a large dataset raise localisation accuracy.

What you can do with it

  • Give natural language instructions to locate a target
  • Work on object, difference and free-form grounding
  • Use direct inference or chained reasoning

Model notes

  • MGrounding-630k holds 630k grounding task samples
  • MIG-Bench results are reported against existing models
  • The model is open sourced on Hugging Face

Pros & cons

✓ What we like

  • Common object, image difference and free-form grounding are covered
  • The model is open sourced on Hugging Face

! What to watch out for

  • Grounding follows the natural language instruction given
  • The material compares results with a 70B model on MIG-Bench

FAQ

What dataset is provided?

MGrounding-630k with 630k multi-image grounding samples.

Where is the model available?

The material states it is open sourced on Hugging Face.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review