PP-OCRv6 release: 50-Language OCR from 1.5M to 34.5M Parameters

PaddlePaddle has released PP-OCRv6, the latest generation of the PaddleOCR universal model family, designed for high-accuracy text detection and recognition across diverse real-world scenarios. This release provides a scalable solution for documents, screenshots, multilingual images, and industrial labels, offering three distinct model tiers to balance accuracy and computational efficiency.

Scalable Model Tiers and Performance

PP-OCRv6 is organized into three tiers to accommodate different deployment constraints, ranging from edge devices to high-performance servers. The medium tier achieves a 86.2% detection Hmean and 83.2% recognition accuracy, representing a significant improvement over the previous PP-OCRv5_server model (+4.6 percentage points in detection and +5.1 percentage points in recognition).

Model	Model size	Detection Hmean	Recognition accuracy	Typical application scenarios
PP-OCRv6_tiny	1.5M params	80.6%	73.5%	Edge devices, lightweight local OCR, latency-sensitive demos, constrained environments
PP-OCRv6_small	7.7M params	84.1%	81.3%	Mobile, desktop, balanced OCR services, multilingual OCR with lower compute cost
PP-OCRv6_medium	34.5M params	86.2%	83.2%	Accuracy-oriented OCR, server-side pipelines, industrial OCR, document ingestion, multilingual OCR

Architectural Improvements

PP-OCRv6 introduces specific architectural updates to both the detection and recognition stages of the OCR pipeline to improve performance on complex, real-world inputs.

Unified Backbone

The model family utilizes PPLCNetV4 as a unified backbone for both text detection and text recognition. This provides architectural consistency across the tiny, small, and medium tiers.

Enhanced Text Detection with RepLKFPN

To address challenges like small, dense, or rotated text, the detection module has been upgraded with RepLKFPN. This is a lightweight large-kernel feature pyramid network designed to improve multi-scale text detection while maintaining inference efficiency.

Improved Recognition with EncoderWithLightSVTR

For the recognition stage, PP-OCRv6 employs EncoderWithLightSVTR. This architecture combines local context modeling with global attention, which enhances recognition quality for challenging text crops, such as multilingual text, screen text, and noisy image regions.

Multilingual Support and Deployment

PP-OCRv6 provides unified multilingual support within the small and medium tiers. These models support 50 languages, including Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin-script languages, reducing the need for multiple specialized models.

Inference Backends

Developers can integrate PP-OCRv6 using several inference backends via the PaddleOCR library:

• Transformers: A Hugging Face / PyTorch-oriented inference path.
• ONNX Runtime: A portable path for ONNX-based deployment environments.
• Paddle Inference: The native Paddle inference format.

Model assets are available on the Hugging Face Hub in multiple formats, including safetensors, Paddle inference models, and ONNX models.