Skip to content

ViTLP

A vision-guided text layout pretraining model

Visit official site

What ViTLP is

The model is described as a vision-guided generative text layout pretraining model that detects and recognises text directly on document images.

What you can do with it

  • Locate and recognise text on a document image
  • Use the 380M parameter ViTLP-medium checkpoint
  • Run batch decoding with the decode.sh script

Application scenarios

  • Extract text from scanned historical documents
  • Automate information extraction from case files
  • Analyse contract pages for key clauses

What to watch out for

  • The medium version balances compute and data limits
  • The stated 5 to 10 seconds is one page on an Nvidia 4090
  • Setup includes installing requirements and model weights

Pros & cons

✓ What we like

  • Pretrained weights are listed on Hugging Face
  • Integration code and instructions are provided

! What to watch out for

  • The 380M version is described as a balance under limited compute
  • The 5 to 10 second figure is for one page on an Nvidia 4090

FAQ

How large is the medium version?

ViTLP-medium is described as having 380M parameters.

How fast is a page?

The material states 5 to 10 seconds for one page on an Nvidia 4090.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review