QIN2DIM/hcaptcha-challenger

What it solves

It provides a way to automatically solve hCaptcha challenges without relying on third-party anti-captcha services or Tampermonkey scripts, using a combination of specialized AI models and multimodal large language models (MLLMs).

How it works

The project uses a pluggable resource system to handle different types of captcha challenges based on the task:

• Image Labeling (Binary/Point): Uses ResNet ONNX for classification and YOLOv8 ONNX for detection.
• Area Selection (Bounding Box): Uses YOLOv8 ONNX segmentation.
• Multiple Choice: Uses ViT ONNX zero-shot motion.
• Drag and Drop: Employs a Spatial Chain-of-Thought approach.
• Advanced Tasks: Integrates CLIP-ViT for self-supervised challenges and AIOps Multimodal LLMs for agentic workflows.

Who it’s for

Developers and automation engineers who need to bypass hCaptcha challenges programmatically using local or AI-driven models.

Highlights

• No Third-Party Dependencies: Operates independently of paid anti-captcha services.
• Multimodal Approach: Combines traditional computer vision models (YOLO, ResNet, ViT) with modern MLLMs.
• Agentic Workflow: Supports complex challenge solving through an AI-driven agentic approach.
• Integrated Pipeline: Includes workflows for data collection, annotation (via Roboflow), and model training/validation via Colab.