inclusionAI/ArmorOCR
ArmorOCR is an 8-billion parameter, two-stage framework developed by inclusionAI for grounded adversarial OCR perception, built upon Qwen3-VL-8B-Instruct. This model specializes in robust optical character recognition, enabling single-pass inference directly on original images without requiring visual transformations or external tools. It is designed to accurately identify text within images, even under adversarial conditions, making it suitable for applications requiring resilient text extraction.
Loading preview...
ArmorOCR: Grounded Adversarial OCR Perception
ArmorOCR is an 8-billion parameter, two-stage framework developed by inclusionAI, specifically designed for grounded adversarial OCR perception. Built on the Qwen3-VL-8B-Instruct base model, ArmorOCR focuses on robustly identifying text within images, even when faced with adversarial inputs.
Key Capabilities
- Single-Pass Inference: Processes original images directly without the need for inference-time visual transformations or tool assistance, streamlining the OCR pipeline.
- Adversarial Robustness: Engineered to perform effectively in challenging scenarios, making it resilient against adversarial attacks on visual text.
- Grounded Perception: Provides a framework for understanding and extracting text with a strong grounding in the visual input.
- Qwen3-VL-8B-Instruct Base: Leverages the capabilities of a powerful vision-language model for enhanced image understanding and text recognition.
Good For
- Applications requiring highly robust optical character recognition.
- Scenarios where text extraction needs to be resilient to image perturbations or adversarial conditions.
- Developers looking for an integrated solution for visual text identification without complex pre-processing steps.
For detailed training information, the AdvSpot benchmark, and evaluation scripts, refer to the GitHub repository. The model is released under the Apache License 2.0, with usage also subject to the base model's license.