PaddleOCR-VL-1.5: Compact & Robust Document Parsing

• PaddleOCR-VL-1.5 is an ultra-compact multi-task vision-language model that integrates distortion-aware layout analysis with diverse document element recognition.
• It employs a Transformer-based pipeline and advanced augmentation techniques to achieve 94.5% accuracy on OmniDocBench v1.5 and superior robustness under real-world distortions.
• The model supports end-to-end document parsing with outputs in Markdown or JSON, enabling efficient extraction of text, tables, formulas, charts, and seals.

PaddleOCR-VL-1.5 is an ultra-compact (0.9 billion parameter) multi-task Vision–LLM (VLM) for end-to-end document parsing, robust layout analysis in unconstrained settings, and recognition of diverse document elements, including text, tables, mathematical formulas, charts, seals, and grounded scene text. Achieving a state-of-the-art (SOTA) accuracy of 94.5% on OmniDocBench v1.5 and surpassing rival VLMs in real-world physical robustness, the model is notable for coupling a unified, distortion-aware layout module with a multitask-enabled VLM, while maintaining high computational efficiency and compactness (Cui et al., 29 Jan 2026).

1. Model Architecture and Pipeline

The PaddleOCR-VL-1.5 pipeline consists of two principal stages: layout analysis and vision–language recognition. Initially, document pages are converted to images and processed by PP-DocLayoutV3, a Transformer-based instance segmentation module using an RT-DETR visual backbone. The system predicts classification labels (e.g., paragraph, table, formula), quadrilateral region masks for non-axis-aligned elements, and absolute reading order with a Global Pointer mechanism. The resulting geometry-rectified regions are then input to PaddleOCR-VL-1.5-0.9B, which recognizes content across six tasks: plain OCR, table (TEDS), formula (LaTeX/markup), chart summarization, seal (stamp text), and grounded four-point text spotting. A post-processing module merges tables, hierarchically refines headings, and emits structured outputs in Markdown or JSON.

Table 1. Core Pipeline Stages

Stage	Module	Functionality
1: Layout	PP-DocLayoutV3	Class/segmentation/reading order
2: Recognition	PaddleOCR-VL-1.5-0.9B	OCR, tables, formulas, charts, seals, spot
Post-processing	Output Normalizer	Table merge, heading structure, export

Notable architectural features include a NaViT-style “Native Resolution” ViT that preserves high spatial detail (up to $1280\times28\times28$ resolution), an ERNIE-4.5-0.3B LLM fused via adaptive MLP, and unified object detection extended with a segmentation head and global reading order via the pairwise precedence score:

$$S_{i,j} = \frac{(W_q q_i)^\top (W_k q_j) - (W_q q_j)^\top (W_k q_i)}{\sqrt{d_h}}, \; S_{i,j} = -S_{j,i}$$

The global voting mechanism sorts elements by $V_j = \sum_{i\neq j}\sigma(S_{i,j})$ to ensure a unique linear reading order.

Each recognition task is handled via a multitask encoder–decoder framework using dedicated instruction prefixes and output heads. The overall training loss is:

$$L_{\rm total} = \lambda_{\rm cls}L_{\rm cls} + \lambda_{\rm box}L_{\rm box} + \lambda_{\rm mask}L_{\rm mask} + \lambda_{\rm ord}L_{\rm order} + \sum_{t\in\{\rm OCR,Tab,Form,Chart,Seal,Spot\}}\lambda_t L_t$$

where constituent terms represent classification, box, mask, and order losses for layout, and sequence generation losses for the content tasks.

2. Datasets, Pretraining, and Data Augmentation

Layout analysis uses a 38,000-page manually annotated corpus spanning academic papers, reports, invoices, comics, and CAD drawings with 25 structure classes and explicit reading order. VLM pretraining utilizes 46 million image–text pairs (up from 29M in v0.9B), balancing scanned and in-scene/photographic documents, multilingual scripts, and explicit seal/text-spotting samples. Instruction fine-tuning covers 5.6 million samples across six tasks, with targeted augmentation for formulas and seals.

Distortion-aware augmentation simulates scanning artifacts, skew, non-rigid warping, moiré, and uneven lighting. Formula samples receive Gaussian lighting and screen-defect effects. Recognition hard-example mining employs Uncertainty-Aware Cluster Sampling (UACS), using CLIP-style embeddings for style clustering and dynamic upsampling of high-variance clusters.

Optimization incorporates AdamW (PP-DocLayoutV3: weight decay $1\mathrm{e}{-4}$, constant LR $2\times10^{-4}$, 32-batch, 150 epochs; VLM: batch 128, max LR $5\times10^{-5}$ pretraining, $8\times10^{-6}$ post-training, all modules jointly trainable).

3. Benchmark Results and Comparative Performance

On OmniDocBench v1.5, PaddleOCR-VL-1.5 establishes new accuracy records with an overall 94.50% (up from 92.86% in v0.9B), including text edit = 0.035, formula CDM = 94.21%, table TEDS = 92.76% (TEDS-S 95.79%), and reading order edit = 0.042. These improvements (+1.64% overall, +2.99% formula CDM) surpass general VLMs such as Qwen3-VL-235B and Gemini-3 Pro.

For robustness, Real5-OmniDocBench benchmarks the model under five real-world physical distortions—scanning, warping, screen-photography, illumination, skew—using one-to-one test sets with preserved ground-truth. PaddleOCR-VL-1.5 achieves 92.05% overall (compared to v0.9B's 85.54% and ≈89% for 235B VLMs).

Table 2. Benchmark Accuracy

Benchmark	v1.5 Accuracy	v0.9B Accuracy	Qwen3-VL-235B/Gemini-3 (approx.)
OmniDocBench v1.5	94.50%	92.86%	<94.50%
Real5-OmniDocBench	92.05%	85.54%	≈89%

By maintaining accuracy and robustness despite its compact parameter count, PaddleOCR-VL-1.5 is distinct among multitask VLMs.

4. Multi-Task Recognition Extensions

The model's multi-task head supports six recognition modalities. Seal recognition is performed using a dedicated autoregressive head and normalized edit distance (NED) as evaluation metric, achieving NED = 0.138 (Qwen3-235B: 0.382) on a 300-image, high-diversity test set incorporating shapes and text deformations. Grounded text spotting models sequences interleaving text with eight location tokens referencing four-point polygons, optimized by sequence cross-entropy. On a proprietary ten-scenario benchmark, the system attains average spotting accuracy of 0.8621 (compared to ≈0.63–0.67 for competing systems).

Instruction-following is further improved through Group Relative Policy Optimization (GRPO), which provides sample-level relative advantages to promote effective instruction alignment and focused reward convergence on challenging samples.

5. Practical Deployment and Efficiency

The inference pipeline is designed for asynchronous, multi-threaded operation with dynamic minibatching from PDF to output. On a single NVIDIA A100 GPU running FastDeploy v2.3, the end-to-end time for all 512 OmniDocBench pages is 944.4s (1.4335 pages/s, 2016.6 tokens/s), with typical GPU memory usage ranging from 40–80GB. Deployment is feasible on widely available accelerators including H800, H20, RTX3060.

Public model weights and code are distributed via GitHub (https://github.com/PaddlePaddle/PaddleOCR) and HuggingFace (https://huggingface.co/PaddlePaddle). The pipeline supports document layout prediction, full document parsing, seal recognition, and text spotting, culminating in markdown-structured or JSON outputs.

6. Technical Significance and Broader Context

PaddleOCR-VL-1.5 demonstrates that highly compact VLMs, when paired with architectural innovations in layout modeling and data-efficient multi-tasking, can achieve and even surpass the accuracy and robustness of much larger general VLMs for document parsing. The explicit modeling of distortion robustness (as evidenced on Real5-OmniDocBench), introduction of new task heads (seal recognition, grounded OCR), and reinforcement-driven instruction alignment illustrate a movement towards unifying document AI capabilities within survey-ready, efficient models.

A plausible implication is that high-performing document parsing in the wild—across scanned, photographed, and physically distorted inputs—can be achieved without recourse to massive generalist VLMs, provided domain-aligned architecture and multitask curricula are implemented. The system’s public accessibility and inference profile suggest immediate applicability in enhanced, real-world document workflows across academic, administrative, and industrial domains (Cui et al., 29 Jan 2026).