ModelCub

The complete computer vision platform. Beautiful, powerful, local-first, free.

Quick Start • Features • Roadmap • Documentation • Discord

---

🎯 Why ModelCub?

You're tired of:

• Roboflow charging $8,000+/year as you scale
• Sending your medical/pharma data to someone else's servers
• Juggling Label Studio + DVC + Ultralytics + custom scripts
• Datasets breaking between experiments
• Annotation tools that feel like they're from 2010

You want:

• ✅ Complete platform (label → train → deploy)
• ✅ Beautiful, modern interface
• ✅ 100% local, zero cloud dependencies
• ✅ Free forever, no limits
• ✅ Git-like version control for datasets
• ✅ Auto-fix broken datasets with one command

ModelCub is what we wish we had when building computer vision systems for medical imaging. No cloud lock-in, no pricing tiers, just a powerful tool that works offline and respects your privacy.

✨ Features

Dataset Operations

• Smart Import: YOLO, Roboflow exports, COCO, or unlabeled images
• Health Scoring: Automatic dataset validation with actionable insights
• Auto-Fix: One command to repair corrupt images, fix bounds, remove duplicates
• Version Control: Git-like commits and visual diffs for datasets
• CLI + SDK: Powerful terminal commands and clean Python API

Annotation (Coming Soon)

• Modern Canvas: Smooth drawing with Konva.js, keyboard shortcuts
• Real-time Validation: Catch issues as you label
• Review Mode: Flag suspicious annotations, batch operations
• Zero Latency: Local processing, 60fps canvas rendering

Training Pipeline (Coming Soon)

• Auto-Optimization: Smart defaults based on your GPU and dataset
• Live Monitoring: WebSocket updates, no page refresh needed
• Multi-Format Export: ONNX, TensorRT, TorchScript, CoreML
• Reproducible: Full experiment tracking and config versioning

Built Different

• API-First: CLI → SDK → Core API. Everything is composable
• Stateless UI: Kill server, restart, nothing lost. All state in filesystem
• No Lock-In: YOLO format internally, export anywhere
• Privacy-First: Zero telemetry, 100% offline capable

🚀 Quick Start

# Install (2 minutes)
pip install modelcub

# Create project
modelcub project init my-cv-project
cd my-cv-project

# Add dataset (supports YOLO, Roboflow, COCO)
modelcub dataset add --source ./data --name bears-v1

# Validate and auto-fix
modelcub dataset validate bears-v1
modelcub dataset fix bears-v1 --auto

# Launch UI
modelcub ui

That's it. Your dataset is imported, validated, and ready.

Python SDK

from modelcub import Project, Dataset

# Initialize
project = Project.init("my-cv-project")

# Load dataset
dataset = Dataset.load("bears-v1")

# Validate
report = dataset.validate()
print(f"Health Score: {report.health_score}/100")

# Auto-fix issues
fix_report = dataset.fix(auto=True)
print(f"Fixed {fix_report.total_fixed} issues")

# Get statistics
stats = dataset.stats()
print(f"Classes: {stats.class_distribution}")

🎨 What Makes ModelCub Special?

1. Auto-Fix That Actually Works

Most tools tell you what's broken. ModelCub fixes it.

$ modelcub dataset fix bears-v1 --auto

🔧 Auto-fixing bears-v1...

✅ Fixed 23 issues:
   • Removed 2 corrupt images
   • Clipped 8 out-of-bounds boxes
   • Removed 4 duplicates
   • Fixed 9 invalid labels

Health Score: 67/100 → 94/100

Backup saved: .modelcub/backups/bears-v1_20250115_143022

2. Dataset Version Control

Git for datasets with visual diffs.

$ modelcub dataset commit -m "Added 200 outdoor images"
$ modelcub dataset diff v1 v2 --visual

📊 Dataset Diff: bears-v1 (v1 → v2)
   Images:  847 → 1,047 (+200)
   Labels:  2,847 → 3,412 (+565)

   Class distribution changed:
   • grizzly: 45% → 38%
   • polar: 35% → 40%
   • black: 20% → 22%

3. Beautiful, Fast UI

Modern interface that doesn't feel like enterprise software from 2010.

• Dark mode by default (light mode available)
• 60fps canvas rendering
• Keyboard-first navigation
• Zero latency on local files

4. Privacy-First Architecture

Your data never leaves your machine.

• ✅ Works 100% offline
• ✅ No telemetry, no tracking
• ✅ No account required
• ✅ HIPAA/GDPR friendly
• ✅ Perfect for medical/pharma/defense

📊 Comparison

	ModelCub	Roboflow	Label Studio + Ultralytics
Annotation	✅	✅	✅
Training	✅	✅	✅
Local-first	✅	❌	✅
Auto-fix	✅	❌	❌
Version control	✅	Basic	Manual
Visual diff	✅	❌	❌
Integrated	✅	✅	❌
Pricing	Free	$500-8k/mo	Free
Setup time	2 min	5 min	30+ min

🗺️ Roadmap

✅ Phase 1: Foundation (Complete)

• Project management
• Dataset import (YOLO, Roboflow, COCO)
• CLI with all core commands
• Python SDK
• FastAPI backend
• React frontend with routing

🚧 Phase 2: Dataset Operations (In Progress)

• Dataset validation with health scoring
• Auto-fix system with backups
• Version control and diff
• Visual diff UI
• Export to multiple formats

📅 Phase 3: Annotation (February 2025)

• Canvas-based annotation tool
• Rectangle and polygon drawing
• Keyboard shortcuts
• Auto-save and undo/redo
• Review mode

📅 Phase 4: Training (March 2025)

• YOLO training integration
• Auto-configuration
• Real-time progress (WebSocket)
• Model evaluation
• Multi-format export

🔮 Future

• Active learning
• Multi-annotator consensus
• Custom augmentation plugins
• Team collaboration features
• Cloud sync (optional)

🎓 Who Is This For?

Indie Developers & Startups

Building CV products without the Roboflow tax. Use the savings to hire engineers.

Medical/Pharma/Defense

HIPAA-compliant, air-gapped training. Your data stays on your servers.

Research Labs

Reproducible experiments with full audit trails. Perfect for papers.

Anyone Fed Up With Cloud Platforms

Own your data. Own your tools. Pay nothing.

💻 Architecture

CLI → Python SDK → Core API ← FastAPI ← React UI
                     ↓
            File System State
         (.modelcub/ directory)

Key Principles:

• API-First: Everything is composable
• Stateless: No hidden database, all state in files
• Format-Agnostic: YOLO internally, import/export anything
• Git-Friendly: Human-readable text files

📚 Documentation

• Installation Guide
• Quick Start Tutorial
• CLI Reference
• Python SDK API
• Architecture Overview

🤝 Contributing

We welcome contributions! ModelCub is built in the open.

Areas we need help:

• Web UI/UX improvements
• Testing and bug reports
• Documentation and tutorials
• Format converters (more import/export formats)
• Example workflows

See CONTRIBUTING.md for guidelines.

🌟 Star History

💬 Community

• Discord: Join our server (coming soon)
• GitHub Discussions: Ask questions, share projects
• Twitter: @ModelCub (coming soon)

📄 License

MIT License - see LICENSE for details.

Built with ❤️ by developers who got tired of expensive cloud platforms.

---

If ModelCub saves you from Roboflow's pricing, star the repo! ⭐

⬆ Back to top