Escaping PaddlePaddle for PyTorch, one OCR model at a time

What it does

PaddleOCR2Pytorch converts Baidu’s PaddleOCR models—detection, direction classification, and text recognition—into PyTorch-compatible weights you can load and run. It covers the PP-OCR family (mobile and server variants), supports multilingual text including Chinese, Japanese, Korean, and European languages, and handles vertical text and long-form recognition. The project also serves as a reference implementation for anyone trying to translate Paddle dynamic-graph code to PyTorch.

The interesting bit

The README opens with the word “白嫖” (roughly: freeloading). The author’s candor matches the project’s purpose: letting you extract value from PaddleOCR’s heavily optimized models—some compressed down to 3.5MB—without committing to the PaddlePaddle ecosystem. It’s a bridge, not a rewrite.

Key highlights

• Pre-converted PyTorch weights available via Baidu Netdisk (extraction code: 6clx)
• Supports PP-OCRv2 through v5, including the latest v5 release with handwritten-text improvements
• Multilingual coverage: 27+ languages in the model zoo
• Includes formula recognition (CAN), text super-resolution (Text Telescope), and layout analysis pipelines
• Mobile-optimized variants for CPU inference

Caveats

• Model downloads require Baidu Netdisk, which is region-restricted and slow outside China
• The TODO list is extensive: PP-OCRv5’s document-orientation modules, PP-StructureV3 PDF parsing, and PP-ChatOCRv4 integration are all marked unfinished
• No training code; inference only

Verdict

Worth a look if you’re already invested in PyTorch and need battle-tested OCR models without ecosystem friction. Skip it if you need end-to-end training or live outside Baidu Netdisk’s effective service area.