clovaai/donut

Donut combines a vision encoder (Swin Transformer backbone) with an autoregressive decoder to jointly perform document image understanding and structured text generation end-to-end. SynthDoG generates multilingual synthetic training data (English, Chinese, Japanese, Korean) to enable flexible pre-training across diverse document types and languages without requiring real annotated datasets. The model integrates seamlessly with Hugging Face's transformers library and achieves state-of-the-art results on document parsing, classification, and VQA tasks with sub-second inference.

6,815 stars. No commits in the last 6 months.