Skip to content

ViTMatte

Image matting built on plain vision transformers

Visit official site

What ViTMatte is

The system is described as an image matting approach that adapts pre-trained plain vision transformers, balancing performance and compute with mixed attention and a convolutional neck.

What you can do with it

  • Add fine detail through a lightweight convolution module
  • Pick a pre-training strategy that suits the case
  • Run matting under different inference settings

Getting started

  • Install the required libraries and tools
  • Download and unpack the code base
  • Select suitable pre-trained model weights
  • Prepare an input image together with its trimap

Pros & cons

✓ What we like

  • Several pre-training strategies aid generalisation
  • The architecture is described as simple to apply

! What to watch out for

  • An input image and a matching trimap are required
  • Pre-trained weights must be chosen before running

FAQ

What is the base model?

The material states pre-trained plain vision transformers.

Which benchmarks are named?

Composition-1k and Distinctions-646 are listed as common matting benchmarks.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review