Research map / Vision-language and multimodal / Multimodal LLMs

PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Presents PP-OCRv5, a 5M-parameter two-stage OCR system whose gains come from a data-centric study of data difficulty, label accuracy, diversity, and quantity, and a 22.6M-sample training set.

Problem

Lightweight OCR models hit a performance plateau and are assumed to lag billion-parameter VLMs, which suffer from coarse localization, hallucination, and high compute cost for OCR tasks.

Approach

Instead of changing architecture, the paper systematically quantifies three training-data dimensions (difficulty via model confidence, label accuracy via synthetic noise, and visual diversity via CLIP features plus K-Means clustering) and scales data accordingly to lift a lightweight two-stage pipeline.

Key results

• PP-OCRv5 reaches 80.1% weighted accuracy on the in-house benchmark, up from 53.0% for PP-OCRv4.
• On OmniDocBench, PP-OCRv5 (5M params) gets 0.067 ALL_avg normalized edit distance, best among specialized OCR tools, but Qwen3-VL (235B) scores 0.026, which is better.
• Label noise from 0% to 20% only reduces accuracy from 0.7188 to 0.7055, and increasing diversity from 200 to 1000 clusters raises accuracy from 0.5860 to 0.6398.

Surprising findings

• The model is robust to label noise, losing only 1.33 points at 20% corruption.
• Very-high-confidence samples (>0.97) hurt performance relative to a mid-confidence sweet spot, so more data is not always better.

Limitations

• The 22.6M training set is in-house and not released, so the main results are hard to reproduce.
• Several headline claims (rivaling billion-parameter VLMs) rest on a table where Qwen3-VL outperforms PP-OCRv5 on the overall metric.
• Ablations use synthetic label noise and a bootstrap model rather than real annotation errors.

Experimental setup

Datasets: in-house 22.6M recognition training set OmniDocBench

Benchmarks: OmniDocBench in-house multi-scenario OCR benchmark

Metrics: normalized edit distance weighted recognition accuracy

Baselines: PP-OCRv4 PP-OCRv3 Qwen3-VL Qwen2-VL InternVL2 GPT-4o Tesseract-OCR EasyOCR Surya Mathpix GOT-OCR

Builds on: PP-OCRv4 DB SVTR GTC CLIP

Closest papers by meaning

• OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.91)
• MME-VideoOCR: Evaluating OCR-Based Capabilities of Multimodal LLMs in Video Scenarios (NeurIPS 2025 · similarity 0.88)
• OCR-Reasoning Benchmark: Unveiling the True Capabilities of MLLMs in Complex Text-Rich Image Reasoning (ICLR 2026 · similarity 0.88)
• Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.87)
• VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• A Picture is Worth More Than 77 Text Tokens: Evaluating CLIP-Style Models on Dense Captions (CVPR 2024 · similarity 0.87)