Parse scanned documents into text, tables, formulas, and charts

PaddleOCR-VL-1.6 is an open document parsing model from PaddlePaddle. It is built for more than reading text from an image. The model targets common document elements such as text blocks, tables, formulas, charts, seals, and spotted text regions.

That matters when you are working with messy document images instead of clean screenshots. A scanned report, research page, table-heavy PDF screenshot, receipt, slide, or form can fail in several ways: text can be recognized but lose reading order, tables can flatten into paragraphs, and formulas can turn into unusable strings. PaddleOCR-VL-1.6 gives you a focused workflow for testing those failure points directly.

The public demo is the fastest way to see the output shape. The backing model card also gives local paths through PaddleOCR CLI, PaddleOCR's Python API, and a transformers example for element-level recognition. Use the hosted Space for a quick first pass, then use the model card when you need a reproducible local setup.

Start with the PaddleOCR-VL-1.6 Online Demo and upload one document image with real structure. A clean one-line sample will not tell you much. Use a table from a report, a screenshot with small UI text, a page with formulas, or a chart-heavy document image.

For the first useful test, run the same image through the available recognition modes. Use full document parsing when the page has mixed content, then try targeted recognition for text, formulas, tables, charts, spotting, or seals when the image has one element you care about most.

Look at three things on the first pass: whether the reading order makes sense, whether structured content stays structured, and whether small or unusual text survives. If the demo result is worth keeping, move to the PaddleOCR model page for local reproduction. The official examples include a CLI command, a Python API path, and an optional inference-server setup for heavier use.

The Hugging Face Space is useful for testing, but its app calls a hosted backend API rather than loading the model directly inside the Space. For a reproducible workflow, treat the model card's PaddleOCR and transformers examples as the local source of truth. Serious local use also expects a proper PaddlePaddle setup and, for best performance, GPU-oriented tooling.

• Extract structured text from scanned pages, receipts, screenshots, and document photos.
• Test whether tables and formulas survive OCR well enough for cleanup or downstream parsing.
• Turn chart-heavy or layout-heavy pages into output you can inspect before building a full ingestion pipeline.
• Compare a document-specific OCR model against a general vision-language model on your own files.
• Use the browser demo as a quick filter before investing in a local PaddleOCR setup.

Try the demo

View model page