Skip to content

Florence-VL

A vision-language model with generative visual encoding

Visit official site

What Florence-VL is

Florence-VL is described as a vision-language model that uses a generative visual encoder and depth-breadth fusion, developed on top of the LLaVA project.

What you can do with it

  • Run pretraining and fine-tuning code
  • Pick a 3B or 8B checkpoint
  • Try the model demo on Hugging Face

Getting started

  • Create a Python virtual environment
  • Download pretraining and instruction data
  • Set training script variables for your paths
  • Launch the training run

Model notes

  • Installation and usage docs are included
  • lmms-eval is supported for evaluation
  • Depth-breadth fusion targets complex tasks

Pros & cons

✓ What we like

  • Installation and usage docs are provided
  • lmms-eval can run multimodal evaluation

! What to watch out for

  • Setup starts with a Python virtual environment
  • Training needs datasets and a training script

FAQ

Which checkpoint sizes?

The material lists 3B and 8B checkpoints.

What is it based on?

The project is described as built on LLaVA.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review