Ink2Pixel

Transform Handwriting into Structured Digital Content with State-of-the-Art Vision Language Models.

Ink2Pixel is a premium document digitization platform that bridges the gap between physical ink and digital pixels. Powered by the Qwen2.5-VL vision-language model, it accurately transcribes complex handwritten notes, mathematical formulas, and structured documents into clean, editable formats.

---

Key Features

• VLM-Powered Transcription: Leverages Qwen2.5-VL for high-fidelity extraction of text and math, even from challenging handwriting.
• Mathematical Excellence: Native support for LaTeX math environments, ensuring formulas are preserved with academic precision.
• Multi-Format Export: Digitized content can be exported as:
  • Markdown (.md)
  • LaTeX (.tex)
  • HTML (.html)
  • JSON (.json)
  • Microsoft Word (.docx)
• Modern FastHTML Interface: A responsive, high-performance web UI designed for speed and clarity.
• PDF Support: Process multi-page documents with automatic page break handling.

---

Quick Start

1. System Requirements

Ink2Pixel runs locally on your machine. For optimal performance with the 7B VLM, we recommend the following:

Component	Minimum	Recommended
GPU	NVIDIA (8GB VRAM)	NVIDIA (12GB+ VRAM)
RAM	16GB	32GB
OS	Windows / macOS / Linux	Windows 11 / macOS 14+

Important

GPU Compatibility:

• NVIDIA: Fully supported with 4-bit hardware acceleration via CUDA.
• Apple Silicon (M1/M2/M3): Supported via MPS, though quantization may vary. 16GB+ Unified Memory recommended.
• AMD: Experimental. If ROCm is not configured, the system will default to CPU mode.
• CPU Fallback: If no compatible GPU is detected, Ink2Pixel will run on your CPU. This is functional but significantly slower (several minutes per page).

2. Prerequisites

• Python 3.9+
• Pandoc (required for .docx conversion)

3. Install Dependencies

pip install -r requirements.txt

4. Launch the Application

Choose the runner for your operating system:

• macOS/Linux: Double-click run_app.sh
• Windows: Double-click run_app.bat

Privacy & Cleanup: Your data never leaves your machine. For extra security, Ink2Pixel automatically deletes all files in the uploads/ and outputs/ folders whenever the application is closed.

Automated Access: Your default web browser will open automatically to http://localhost:8000 once the server is ready.

---

Project Structure

• app.py: The main FastHTML application and web interface.
• vlm/document_digitizer.py: The core engine handling model loading and inference.
• requirements.txt: Project dependencies.
• legacy/: Historical preprocessing tools and experiments (kept for reference).

---

Technology Stack

• Core: Python
• Frontend: FastHTML & HTMX
• VLM: Qwen2.5-VL-7B-Instruct
• Deep Learning: PyTorch & HuggingFace Transformers
• Optimization: BitsAndBytes (4-bit quantization)
• Document Handling: PyMuPDF & Pandoc

---

License

This project is intended for research and personal digitization. Please refer to the Qwen2.5-VL model license for usage terms related to the underlying VLM.

---

Ink2Pixel — Bridging the analog-digital divide.