donut-base

Extracts text and structured information from document images using a vision-encoder-decoder architecture that combines a CNN-based image encoder with a transformer decoder. The model processes document layouts end-to-end without requiring OCR preprocessing, learning to recognize both text content and spatial relationships. It uses a sequence-to-sequence approach where the encoder converts images to visual embeddings and the decoder generates structured text outputs (JSON, key-value pairs, or markdown) conditioned on the visual context.

Converts document images into dense visual embeddings using a CNN-based encoder (typically ResNet or similar backbone) that extracts spatial and semantic features from the image. The encoder processes the full image in a single forward pass, producing a sequence of patch embeddings or feature maps that capture document structure, text regions, and layout information. These embeddings serve as the input representation for downstream sequence generation or classification tasks.

Generates text sequences conditioned on visual embeddings using a transformer decoder that attends to the encoded image representation. The decoder uses cross-attention mechanisms to align generated tokens with relevant image regions, enabling it to produce coherent text that reflects the document's content and structure. The generation process supports both greedy decoding and beam search, allowing trade-offs between speed and output quality.

Processes multiple document images efficiently through dynamic batching, where the model groups images of similar sizes to minimize padding overhead and maximize GPU utilization. The implementation handles variable-sized inputs by padding to the largest image in each batch, then processes all images in parallel through the encoder-decoder pipeline. Supports both synchronous batch processing and asynchronous queuing for high-throughput scenarios.

Supports fine-tuning the pre-trained model on custom document datasets to adapt it to specific domains (e.g., medical forms, invoices, contracts). The fine-tuning process updates both encoder and decoder weights using supervised learning on labeled document-text pairs. Implements standard training loops with gradient accumulation, mixed precision training, and learning rate scheduling to optimize convergence on domain-specific data.

Supports document understanding across multiple languages (primarily English and Korean, with limited support for other languages) through language-specific decoding strategies. The model's tokenizer and decoder are trained on multilingual text, enabling it to generate output in the language of the input document. Language detection can be performed on input images or specified explicitly to optimize decoding.