PixelGPT

<a href="https://2024.emnlp.org/" target="_blank"> <img alt="EMNLP 2024" src="https://img.shields.io/badge/Proceedings-EMNLP2024-red" /> </a>

This repository contains the official checkpoint for PixelGPT, as presented in the paper Autoregressive Pre-Training on Pixels and Texts (EMNLP 2024). For detailed instructions on how to use the model, please visit our GitHub page.

Model Description

PixelGPT is an autoregressive language model pre-trained exclusively on pixel data using a next patch prediction objective. By processing documents as visual data (pixels), the model learns to predict the next image patch in a sequence, enabling it to handle visually complex tasks without relying on tokenized text. This tokenization-free approach allows PixelGPT to process and understand text rendered as images.

Citation

@misc{chai2024autoregressivepretrainingpixelstexts,
  title = {Autoregressive Pre-Training on Pixels and Texts},
  author = {Chai, Yekun and Liu, Qingyi and Xiao, Jingwu and Wang, Shuohuan and Sun, Yu and Wu, Hua},
  year = {2024},
  eprint = {2404.10710},
  archiveprefix = {arXiv},
  primaryclass = {cs.CL},
  url = {https://arxiv.org/abs/2404.10710},
}