jonstreeter-ComfyUI-Reference-Based-Video-Colorization/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.0.0] - 2025-11-13

🎉 Major Release - Dual Implementation

This release includes a complete rewrite with two colorization methods available side-by-side!

Added

ColorMNet Implementation (New)

• ✨ ColorMNet Video Colorization Node - Modern memory-based temporal colorization
• ✨ ColorMNet Image Colorization Node - Single image colorization
• 🚀 DINOv2-based feature extraction for superior quality
• ⚙️ Multiple memory modes (balanced, low_memory, high_quality)
• ⚡ FP16 mixed precision support for faster processing
• 📊 Performance reports with timing and FPS metrics
• 🎯 Automatic model download (~500MB DINOv2 model)
• 📈 Real-time progress bars in ComfyUI interface

Deep Exemplar Implementation (Enhanced Original)

• ✨ Deep Exemplar Video Colorization Node - Original CVPR 2019 method
• ✨ Deep Exemplar Image Colorization Node - Classic exemplar-based approach
• 📊 Performance reports with timing and FPS metrics
• 📈 Real-time progress bars in ComfyUI interface
• 🎯 Automatic model download for all checkpoints
• 🔧 Enhanced WLS filtering with configurable parameters

Infrastructure

• 📦 Automatic checkpoint downloading for all models
• 🔄 Lazy loading of VGG19 models to prevent import-time errors
• 🐛 Fixed relative import issues across all modules
• 📁 Proper extraction of Deep Exemplar checkpoint ZIP files
• 🎨 Example workflow with both methods for comparison
• 📸 Workflow screenshot for README
• 📚 Comprehensive documentation updates

Fixed

• 🐛 ColorMNet greyscale output issue - proper RGB reference passing
• 🐛 ColorMNet color saturation - normalized reference ab channels
• 🐛 Progress bar not working in ColorMNet nodes
• 🐛 Import errors with absolute vs relative paths
• 🐛 VGG19 model loading at module import time
• 🐛 Deep Exemplar checkpoint extraction path issues
• 🐛 Missing auto-download for Deep Exemplar models

Changed

• 📝 Complete README rewrite with both implementations
• 🏗️ Reorganized project structure for clarity
• 🎯 Node display names clarified: "(New)" for ColorMNet, "(Original)" for Deep Exemplar
• 🔧 Improved error messages and logging
• ⚡ Optimized memory usage and performance

Technical Details

ColorMNet Fixes

1. Reference Image Processing - Changed from grayscale (L×3) to full RGB
2. Color Space Normalization - Proper LAB ab channel normalization (÷110)
3. Output Denormalization - Correct scaling from [-1,1] to LAB range (×110)
4. Progress Updates - Added set_progress() calls for UI updates

Deep Exemplar Improvements

1. Auto-download System - Downloads from GitHub releases automatically
2. ZIP Extraction - Proper handling of nested directory structure
3. VGG Loading - Lazy initialization with correct absolute paths
4. Import Fixes - All relative imports (. and ..) instead of absolute

Performance Reports

All four nodes now output optional STRING reports containing:

• Processing time and FPS
• Frame count and resolution
• Configuration parameters used
• Enable/disable status of features

Performance

Example Benchmarks (768x432, 240 frames):

• ColorMNet (balanced, FP16): ~5.3 FPS, 45s total
• Deep Exemplar (half-res, propagation): ~4.6 FPS, 52s total

[1.0.0] - 2024-XX-XX

Initial Release

• Initial Deep Exemplar implementation
• Basic video and image colorization nodes
• Manual model download required
• WLS filtering support

---

Migration Guide

From 1.x to 2.0

Good News: Both old and new implementations are available!

Changes:

1. Node Names: Original nodes renamed to "Deep Exemplar ... (Original)"
2. New Nodes: "ColorMNet ... (New)" nodes added
3. Outputs: All nodes now have optional performance_report output
4. Auto-download: Models download automatically on first use

Existing Workflows:

• Will continue to work with Deep Exemplar nodes
• Consider trying ColorMNet for comparison
• Performance reports are optional (can be ignored)

Recommended Actions:

1. Update ComfyUI custom node
2. Try the new example workflow
3. Compare ColorMNet vs Deep Exemplar quality
4. Review performance reports to optimize settings

---

Future Roadmap

Planned Features

• Batch processing optimization
• Memory-efficient streaming mode
• Additional colorization algorithms
• Color palette transfer utilities
• Integration with temporal smoothing nodes

Under Consideration

• Web UI for reference image search
• Automatic reference selection
• Multi-reference blending
• Training script for custom models
• Real-time preview mode

---

Version Support

Version	Status	Support Until	Notes
2.0.x	✅ Active	Ongoing	Current release
1.0.x	⚠️ Legacy	2025-12-31	Security fixes only

---

Breaking Changes

2.0.0

• Node display names changed (internal IDs unchanged)
• New output socket added (backwards compatible)
• Model paths changed (auto-downloaded)

Migration: Update nodes in workflow to see new names. Existing connections work unchanged.

---

Contributors

Thanks to everyone who contributed to this release!

• ColorMNet integration and bugfixes
• Deep Exemplar enhancements
• Documentation improvements
• Testing and feedback

---

For detailed technical changes, see the commit history.