Visual Document Understanding
Donut
Decode page images into task-specific fields or answers without a separate OCR stage.
At a glance
- Input
- Page images
- Output
- Field sequences, JSON
- Provenance
- No default field coordinates
- Scanned pages
- Image-based input
- Deployment
- Self-hosted
- Availability
- Open source
- Costs to plan for
- Compute and training costs
Limitations to check
Single-page understanding is not whole-document reasoning. Validate or fine-tune for domain changes and the desired output schema.
Build a meaningful sample
Include your hardest pages: low-quality scans, missing fields, dense tables, and evidence split across pages. Validate output correctness separately from citation accuracy.