AI, Fix My Messy Whiteboard!

A local AI utility for cleaning up whiteboard photographs using image-to-image processing models via Replicate.

Overview

Whiteboard Processor transforms messy whiteboard photos into clean, professional diagrams. It uses various image generation models through Replicate's API to redraw whiteboard content with:

• Clean white backgrounds (no glare or reflections)
• Legible, clear text
• Professional lines, arrows, and shapes
• Preserved layout and spatial relationships

Features

• CLI-based workflow - Drop images into a queue folder and process them interactively
• Multiple model support - Switch between different image models (FLUX, Qwen, Google Nano-Banana)
• Evaluation framework - Compare model outputs side-by-side to find the best results
• Batch processing - Process multiple whiteboard images in one run

Installation

cd app
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Create a .env file in the project root with your Replicate API key:

REPLICATE_API_KEY=your_key_here

Usage

# Run the processor
./run.sh

# Or directly with Python
cd app && python process_whiteboard.py

# Specify a model
python process_whiteboard.py -m nano-banana-pro

# List available models
python process_whiteboard.py --list-models

Place whiteboard images in images/originals/queue/<folder-name>/ and run the processor. Enhanced outputs are saved to images/enhanced/.

Models

The tool currently includes configurations for:

• nano-banana - Google Gemini 2.5 Flash (default, fast)
• nano-banana-pro - Google Gemini 3 Pro (highest quality)
• flux-kontext-pro - FLUX Kontext Pro
• flux-1.1-pro - FLUX 1.1 Pro
• qwen-image-edit - Qwen Image Edit
• qwen-image-edit-plus - Qwen Image Edit Plus

Note: Image generation models evolve rapidly. The current model selection will likely become outdated. Swap in newer/better models as they become available on Replicate by editing the MODELS dict in app/process_whiteboard.py.

Evaluation

An evaluation script is included to compare model outputs:

cd eval
python run_eval.py

Place test images in eval/source/ and the script will run them through all configured models, saving results to timestamped folders in eval/runs/.

---

Replicate Comparison (Dec 02, 2025)

Original whiteboard with text annotations:

Flux Pro 1.1

Pseudotext: high Misinterpretation of system prompt

---

Flux 1.1 Kontext

No pseudotext - excellent replication of handwriting

---

Nano Banana

Pseudotext: Clawpe could have been inferred around. Followed colorisation instruction.

---

Nano Banana Pro

Best Overall

Inferred Claude from Clawpe (non-Pro version didn't!)

---

Qwen Image Edit

Highest level of pseudotext in any of the models

---

Qwen Image Edit Plus

---

License

MIT