Exciting news for developers working with Optical Character Recognition (OCR) and document parsing! PaddleOCR 3.5 marks a significant leap forward, bringing its robust capabilities closer to the widely adopted Hugging Face ecosystem. This latest release introduces the ability to run supported PaddleOCR models using Hugging Face Transformers as an inference backend, offering unprecedented flexibility and integration.
By simply setting engine="transformers", developers can now leverage their familiar Transformers environment to power PaddleOCR’s advanced model series, including PP-OCRv5 for OCR and PaddleOCR-VL 1.5 for sophisticated document parsing. This means you get the best of both worlds: PaddleOCR’s cutting-edge AI for document intelligence, seamlessly integrated into your Transformers-centric workflows. Ready to see it in action? A live demo is available on Hugging Face Spaces.
Unlocking New Possibilities with a Flexible Backend
PaddleOCR 3.5 introduces a much more flexible inference-engine interface, putting developers in control. You can now select your preferred backend via the engine parameter and fine-tune backend-specific options through engine_config. This design empowers you to tailor the deployment to your specific infrastructure and performance needs.
It’s crucial to understand that this update focuses on the inference backend layer, not on replacing PaddleOCR’s core functionalities. PaddleOCR continues to deliver industry-leading OCR and document parsing capabilities. What changes is how these powerful models can be executed, now offering a native option that perfectly fits into Hugging Face-centered development environments.
Why This Integration is a Game-Changer for Document AI
For anyone building applications around RAG (Retrieval Augmented Generation), Document AI, or intelligent document agents, the real challenge often begins long before engaging with a Large Language Model (LLM). The initial hurdle is transforming complex, unstructured document formats into reliable, structured data. This includes everything from PDFs and scanned documents to screenshots, tables, charts, and even intricate formulas.
If this crucial ingestion step is weak or unreliable, downstream LLM workflows can suffer significantly. They might miss vital information, retrieve incorrect context, or produce untrustworthy answers. This is precisely where PaddleOCR shines, offering robust solutions for this challenging document ingestion phase through its advanced OCR and document parsing models.
With PaddleOCR 3.5, connecting these powerful capabilities with your existing Transformers-centered stacks has never been easier. Supported PaddleOCR models can now run with a Transformers backend, allowing PaddleOCR to manage the intricate OCR or document parsing pipeline behind the scenes. For developers, this translates into tangible benefits:
- Reduced Integration Friction: Seamlessly incorporate advanced document processing into your existing ML pipelines.
- Natural Workflow Path: Create a more intuitive journey from raw documents to sophisticated downstream applications like RAG, intelligent agents, search, analytics, or automation.
- Leverage Existing Infrastructure: Utilize your established PyTorch and Transformers setup for model loading, experimentation, and deployment.
Getting Started: A Seamless Transition
Diving into PaddleOCR 3.5 with the Transformers backend is straightforward. First, ensure you have a compatible PyTorch build for your hardware, then install the necessary libraries. For CUDA-enabled GPUs, an example installation might look like this:
python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
python -m pip install "paddleocr==3.5.0" "paddlex==3.5.2" "transformers>=5.4.0"Remember to adjust your PyTorch installation command if you are using a CPU, ROCm, or another specific environment to match your target hardware.
Once installed, you can leverage the Transformers backend with minimal code changes. Here’s how you might run an OCR task using the command-line interface (CLI):
paddleocr ocr \
-i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png \
--device gpu:0 \
--engine transformersAlternatively, for more programmatic control within your Python applications, you can use the PaddleOCR API:
from paddleocr import PaddleOCR
pipeline = PaddleOCR(
device="gpu:0",
engine="transformers",
use_doc_orientation_classify=False,
use_doc_unwarping=False,
use_textline_orientation=False,
engine_config={"dtype": "float32"},
)
results = pipeline.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png")
for result in results:
print(result)The Hugging Face Space demo typically uses float32 for broad compatibility. However, for optimized performance on your own hardware, you can fine-tune backend-specific options through the engine_config parameter. For example, you might specify bfloat16 for improved efficiency where supported:
engine_config = {
"dtype": "bfloat16",
"device_type": "gpu",
"device_id": 0,
"attn_implementation": "sdpa",
}The optimal configuration will naturally depend on your specific model, hardware, and deployment environment. Experimentation with these settings can yield significant performance benefits tailored to your setup.
When to Choose the Transformers Backend
The Transformers backend is particularly beneficial when you want PaddleOCR’s powerful OCR and document parsing capabilities to integrate seamlessly into a Hugging Face-centered stack. This choice is ideal if your current workflow for building RAG, Document AI, search, analytics, or agent applications already heavily relies on PyTorch and Transformers infrastructure for model loading, experimentation, or deployment.
Conversely, for scenarios where maximizing raw OCR or document parsing throughput is the absolute priority, PaddleOCR’s default paddle_static backend is typically the recommended choice. This release isn’t about replacing one backend with another; it’s about providing developers with more strategic options. The goal is to empower you to leverage PaddleOCR’s advanced document intelligence while selecting the inference backend that best aligns with your existing technology stack and performance objectives.
Experience the power and flexibility firsthand! We invite you to try the PaddleOCR 3.5 Transformers demo on Hugging Face Spaces today. You can also explore the PaddlePaddle organization on Hugging Face for more models and resources.
PaddleOCR 3.5 ultimately brings its advanced OCR and document parsing capabilities closer to Transformers-centered workflows, granting developers the freedom and flexibility to construct sophisticated Document AI applications around them.
We extend our sincere gratitude to the dedicated Hugging Face engineers who provided invaluable support throughout the PaddleOCR 3.5 Transformers integration. Special thanks go to Anton Vlasjuk for his comprehensive involvement, including reviewing and merging all related pull requests, ensuring a smooth and successful integration.
We also deeply appreciate the insightful PR reviews and feedback from Raushan Turganbay and Yoni Gozlan. Their expert guidance significantly enhanced the integration quality, refined our documentation, and ultimately improved the developer experience for the wider Hugging Face community. We couldn’t have done it without their contributions!
Source: Hugging Face Blog
