ArmorOCR is a two-stage framework for grounded adversarial OCR perception from Ant Group's inclusionAI, built on Qwen3-VL-8B-Instruct and introduced in the paper "ArmorOCR: Grounded Adversarial Visual Perception via Observation-Transferred Self-Distillation." Its key design goal is robustness to adversarial or degraded visual conditions while enabling single-pass inference directly on the original image, without requiring any inference-time visual transformations (such as cropping, upscaling, or denoising) or external tool assistance. The model follows a structured reasoning-then-answer output format, wrapping its analysis in
<analyze></analyze> tags and the final recognized text in
<answer></answer> tags, and is loaded via the standard
Qwen3VLForConditionalGeneration class in Transformers. Training details, the accompanying AdvSpot adversarial OCR benchmark, and evaluation scripts are hosted separately on the project's GitHub repository, and the model is released under Apache License 2.0, with usage additionally subject to the base Qwen3-VL-8B-Instruct model's license and acceptable-use policy.
