Skip to content

PixelLLM

A vision-language model that aligns words with pixels

Visit official site

What PixelLLM is

The model is described as a vision-language model for image grounding tasks, linking descriptive text with pixel positions.

What you can do with it

  • Generate descriptive text from an input position
  • Generate pixel coordinates from input text
  • Handle dense object description

Model notes

  • Pretraining uses the Localized Narrative dataset
  • Word and pixel alignment is learned in pretraining
  • Results are described on RefCOCO and Visual Genome

Pros & cons

✓ What we like

  • Both text-from-position and position-from-text directions are covered
  • Word-to-pixel alignment is learned in pretraining

! What to watch out for

  • Tasks are centred on image grounding
  • Named datasets are RefCOCO and Visual Genome

FAQ

What does it align?

The material describes learning alignment between words and image pixels on Localized Narrative.

Which tasks are listed?

The feature list names location-conditioned description, referring grounding and dense object description.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review