Skip to content

LLaVA-Mini

One visual token for efficient image and video understanding

Visit official site

What LLaVA-Mini is

The model is described as a multimodal model from the ictnlp team that uses one visual token to understand images and video efficiently and is released as open source.

What you can do with it

  • Analyse video content and understand events and objects quickly
  • Recognise text and objects in images with high efficiency
  • Process long video material of up to three hours
  • Improve the efficiency of image and video understanding

Performance notes

  • Computation is reported to fall by 77 percent
  • Response latency is stated at 40 ms
  • One-token performance is comparable to LLaVA-v1.5
  • Over 10000 frames run on a GPU with 24 GB memory

Getting started

  • Download the LLaVA-Mini model from Hugging Face
  • Run the controller startup script
  • Build the API for the model
  • Launch the interactive interface

Pros & cons

✓ What we like

  • Performance with one token is comparable to LLaVA-v1.5
  • Memory use is greatly reduced for long video work

! What to watch out for

  • The model must be downloaded from Hugging Face before local use
  • Running it needs a controller script, an API and an interactive interface

FAQ

How many visual tokens are used?

The material states an image is represented with one visual token.

What hardware is mentioned?

A GPU with 24 GB memory is stated to handle over 10000 video frames.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review