--- title: "paddleocr" type: "AI Tool" url: "https://aidemos.com/tools/paddleocr" description: "We tested PaddleOCR PP-StructureV3 on mixed PDFs: it kept charts, photos, signatures, and many table values, but lost words, digits, and structure." category: "productivity" website: "https://github.com/PaddlePaddle/PaddleOCR" published: "2026-08-13T18:16:23.890718+00:00" updated: "2026-08-13T18:20:38.078669+00:00" --- # paddleocr Open-source PDF-to-markdown conversion that keeps images and tables in place, but still corrupts text and complex structure on harder documents. ## TL;DR Verdict **Strong layout retention, but not clean enough to trust without review** **Where it wins:** - You need a self-hosted open-source PDF-to-markdown pipeline for mixed documents with text, tables, charts, and images. - You can tolerate some post-processing when tables are complex or when native-digital text is not perfectly preserved. - You want extracted photos, seals, signatures, and charts to stay embedded as real image files in the output. **Main limitation:** You need near-perfect paragraph OCR from clean digital PDFs with no cleanup. **Pricing:** Open source Free `Open source` · `PDF to markdown` · `OCR + layout parsing` · `Retains images as files` **Website:** [Visit paddleocr](https://github.com/PaddlePaddle/PaddleOCR) > **Strong layout retention, but not clean enough to trust without review** > > PaddleOCR’s PP-StructureV3 is a capable self-hosted PDF-to-markdown pipeline: it reliably exports mixed PDFs, keeps charts/photos/signatures as real image files, and preserves many table values. But body text can lose words or digits, wide or rowspan-heavy tables break structurally, headings are inconsistent, and some logos are dropped entirely. It’s useful when you can tolerate cleanup, not when you need near-perfect markdown out of the box. ## Demo Recording [Video: paddleocr demo recording](https://cdn.futuresmart.ai/public/aidemos/ec91ca8cbfbc4e0685af13a1f6bb3385.mp4?v=1) *Video — Google Colab walkthrough of the PaddleOCR PDF-to-markdown workflow.* ## Feature-by-Feature Breakdown ### OCR text extraction from mixed PDF pages **Verdict:** mixed Extracts body text from both native-digital and scanned PDF pages into markdown. In the hybrid earnings report, table-heavy financial report, and scanned research paper, it could recover text but sometimes lost words, letters, or digits depending on document type. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** Strong on clean scanned text, but native-digital paragraphs can still lose words or digits, so the OCR layer is not uniformly trustworthy. ### Table reconstruction **Verdict:** mixed Converts financial tables into markdown/HTML tables and usually preserves labels and values. The exercised inputs include small tables, larger wrapped-label tables, rowspan headers, and side-by-side layouts where structural fidelity becomes fragile. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** Good at reading table content, but structural fidelity breaks down on complex tables long before the data values themselves do. ### Embedded visual retention **Verdict:** mixed Retains non-text visual elements from PDFs as real cropped image files placed near their original location. The cards exercise this on charts/figures as well as photos, signatures, seals, and logos, with some cases collapsing or disappearing entirely. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** It reliably keeps chart regions inline, but the output is image-only rather than data-aware. ### Heading and document hierarchy export **Verdict:** mixed Turns section titles into markdown headings to approximate document structure. The tested pages include section titles from reports and a scanned paper, but heading levels and section flattening can be inconsistent. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** Useful for rough sectioning, but the heading tree is inconsistent enough that a clean TOC or outline needs manual review. ### Reading-order preservation **Verdict:** mixed Keeps content in front-to-back order for most single-column pages and documents. The exercised multi-column layouts show that unrelated blocks can interleave and paragraph boundaries can break at column transitions. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** Normal linear pages are fine, but multi-column layouts and column breaks still need cleanup. ### Caption and footnote anchoring **Verdict:** preserved Keeps captions, unit notes, and footnotes attached to the table or figure they describe, even when the underlying object is degraded. The tested cases include captions under tables and figures where the explanatory text still stays in place. **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Input:** source > **Image** — source **Output:** > **Image** **Bottom line:** Captions and footnotes are generally anchored correctly, which helps even when the table or chart itself is imperfect. ## Open source, no per-page fees Apache 2.0 licensing; compute is the only real cost. | Plan | Price | Notes | | --- | --- | --- | | Open source ★ (tested) | Free | Self-hostable; no per-document billing. | *PaddleOCR / PP-StructureV3 is open source under Apache 2.0, with no usage-based API pricing. The report’s runs were done on Google Colab T4 GPU compute, so the practical cost is infrastructure rather than vendor fees.* ## Is It Right For You? **Use it if** - You need a self-hosted open-source PDF-to-markdown pipeline for mixed documents with text, tables, charts, and images. - You can tolerate some post-processing when tables are complex or when native-digital text is not perfectly preserved. - You want extracted photos, seals, signatures, and charts to stay embedded as real image files in the output. **Skip it if** - You need near-perfect paragraph OCR from clean digital PDFs with no cleanup. - You need complex table structure, rowspan headers, and multi-column layouts to be structurally exact on the first pass. - You need charts converted into usable data rather than preserved only as images. ## Classification - **Category:** productivity - **Subcategory:** documentation-tools - **Type:** text - **Built for:** Other ## Frequently Asked Questions **Q: Does PaddleOCR keep charts in the markdown output?** Yes, it keeps chart regions as real inline image files and places them near the source location. In the report, charts were retained, but no chart data, axis labels, or trends were reconstructed as text. **Q: How well does it handle scanned PDFs?** On the scanned research paper, it performed better than on the native-digital PDFs: the checked paragraph preserved the words correctly, and the main defect was dropped spacing rather than lost content. It still did not recover chart data. **Q: Does it preserve complex tables?** It preserves many table values well, but the structure can break on wrapped labels, rowspan-based headers, short total rows, and side-by-side layouts. Small tables were often exact, while larger tables needed cleanup. **Q: Are images and logos retained?** Many photos, signatures, seals, and figures were retained as real image tags, but some logos were dropped entirely with no placeholder. Image retention is helpful, but not reliable enough to assume nothing was lost. **Q: Is PaddleOCR free to use?** Yes. The report identifies it as open source under Apache 2.0, with no per-page or per-document API fees. The practical cost is just the compute you run it on. ## Similar Tools AI tools similar to paddleocr: - [Docling](https://aidemos.com/tools/docling) — Open-source PDF-to-markdown conversion that is strong on text, headings, and standard tables, but drops charts and other visual assets. - [PyMuPDF4LLM](https://aidemos.com/tools/pymupdf4llm) — Open-source PDF-to-markdown for clean native PDFs, but unreliable on scans, dense tables, and images. - [LiteParse](https://aidemos.com/tools/liteparse) — Open-source PDF-to-markdown parsing that works well on native-digital reports, but degrades on scans, tables, and charts. - [doc2mark](https://aidemos.com/tools/doc2mark) — Open-source PDF-to-markdown that preserves native text and headings, but still struggles with tables, charts, images, and scans. - [MarkItDown](https://aidemos.com/tools/markitdown) — Fast native-PDF text extraction for markdown, but structure, charts, images, and scans are unreliable. - [MinerU](https://aidemos.com/tools/mineru) — Open-source PDF-to-markdown that keeps figures and charts in place, but tables and punctuation can get messy on complex files. ## Need a custom AI solution for this use case? If you are looking to build a custom OCR extraction, PDF-to-markdown conversion, or document parsing workflow for your business or internal workflow, email us at [contact@futuresmart.ai](mailto:contact@futuresmart.ai). ### Found something inaccurate or missing? We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at [collaborate@aidemos.com](mailto:collaborate@aidemos.com).