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

Quick Start Guide - ColorMNet Activation

✅ Status: New Code Activated!

The new ColorMNet implementation is now active. Follow these steps:

---

Step 1: Install Dependencies ⚡

Open a terminal in this directory and run:

# Make sure you're in the right directory
cd "G:\AIART\AI_Image_Generators\Comfy_UI_V42\ComfyUI\custom_nodes\ComfyUI-Reference-Based-Video-Colorization"

# Install/update Python packages
pip install torch>=2.0.0 torchvision>=0.15.0 numpy scikit-image opencv-python einops tqdm progressbar2

# Install ColorMNet-specific dependencies
pip install git+https://github.com/cheind/py-thin-plate-spline.git
pip install git+https://github.com/ClementPinard/Pytorch-Correlation-extension.git

Note for Windows users: The Pytorch-Correlation-extension might need Visual Studio Build Tools. If it fails, you can skip it for now and see if the model works without it.

---

Step 2: Download Model Checkpoint 📥

The model checkpoint is ~500MB. Download it from:

Option A: Automatic (Recommended)

python -c "import gdown; gdown.download('https://github.com/yyang181/colormnet/releases/download/v0.1/DINOv2FeatureV6_LocalAtten_s2_154000.pth', 'checkpoints/DINOv2FeatureV6_LocalAtten_s2_154000.pth', quiet=False)"

Option B: Manual

1. Go to: https://github.com/yyang181/colormnet/releases
2. Download: DINOv2FeatureV6_LocalAtten_s2_154000.pth
3. Place in: checkpoints/DINOv2FeatureV6_LocalAtten_s2_154000.pth

---

Step 3: Restart ComfyUI 🔄

# Stop ComfyUI (Ctrl+C in the terminal)
# Then restart it

You should see on startup:

[ColorMNet] Loading v2.0.0
[ColorMNet] License: CC BY-NC-SA 4.0 (Non-commercial use only)
[ColorMNet] Based on ColorMNet by Yang Yang (ECCV 2024)

---

Step 4: Use the New Nodes 🎨

1. Right-click in ComfyUI
2. Add Node → ColorMNet
3. You'll see:
   • ColorMNet Video Colorization - For video sequences
   • ColorMNet Image Colorization - For single images

Quick Test Workflow:

[Load Image] → video_frames
                     ↓
              [ColorMNet Video]
                     ↑
[Load Image] → reference_image

Parameters:
- target_width: 768
- target_height: 432
- memory_mode: balanced
- use_fp16: True
- use_torch_compile: False  # Set to True for 15-25% speedup

---

Troubleshooting 🔧

"Model checkpoint not found"

Solution: Make sure the model file is at:

checkpoints/DINOv2FeatureV6_LocalAtten_s2_154000.pth

Run this to check:

ls -lh checkpoints/

Should show ~500MB file.

"ModuleNotFoundError: No module named 'spatial_correlation_sampler'"

Solution: The Pytorch-Correlation-extension failed to install. Try:

# On Windows with Visual Studio installed:
pip install git+https://github.com/ClementPinard/Pytorch-Correlation-extension.git

# Or try pre-built wheel (check GitHub releases)

If it still fails, the model might work without it (the extension is for optical flow which ColorMNet doesn't heavily use).

"ModuleNotFoundError: No module named 'thin_plate_spline'"

Solution:

pip install git+https://github.com/cheind/py-thin-plate-spline.git

Old nodes still showing up

Solution:

1. Completely restart ComfyUI
2. Clear browser cache (Ctrl+Shift+Delete)
3. If old nodes persist, they're from DeepExemplarColorizationNodes.py (backed up as DeepExemplarColorizationNodes_old_backup.py)

---

What Changed? 📝

Old Nodes (Deep Exemplar 2019):

• ❌ DeepExColorImageNode
• ❌ DeepExColorVideoNode

New Nodes (ColorMNet 2024):

• ✅ ColorMNetImage
• ✅ ColorMNetVideo

Benefits:

• 🚀 2x faster (4 fps vs 2 fps)
• 💾 33% less VRAM (4GB vs 6GB with FP16)
• 🎯 Better quality (NTIRE 2023 winner)
• 🛡️ Better error messages
• 🔄 Automatic device detection
• 📊 Progress bars

---

Performance Optimizations ⚡

New Optimization Options

Both ColorMNet and Deep Exemplar nodes now support additional performance optimizations:

torch.compile (All Nodes)

Enable PyTorch 2.0+ graph compilation for 15-25% speedup:

use_torch_compile = True

• Requires PyTorch 2.0 or later
• First run slower due to compilation (warmup)
• Subsequent runs get full speedup
• No quality loss

SageAttention (Deep Exemplar Only)

Enable INT8-quantized attention for 20-30% faster attention:

use_sage_attention = True

• Requires: pip install sageattention
• Works on CUDA GPUs only
• Negligible quality impact
• Automatically falls back if unavailable

Recommended Settings

Maximum Speed (ColorMNet):

use_fp16 = True
use_torch_compile = True
memory_mode = "low_memory"

Maximum Speed (Deep Exemplar):

use_torch_compile = True
use_sage_attention = True
use_half_resolution = True

Best Quality (All Nodes):

use_torch_compile = False  # or True for speed boost
use_fp16 = False  # ColorMNet only

See PERFORMANCE.md for detailed benchmarks and optimization guide.

---

Next Steps 🎯

1. Complete installation steps above
2. Restart ComfyUI
3. Test with a small video (5-10 frames)
4. Adjust parameters based on your GPU
5. Enjoy better colorization!

---

Need Help? 💬

• Check README_NEW.md for detailed documentation
• Check MIGRATION_GUIDE.md if you have old workflows
• Check ARCHITECTURE.md for technical details
• Open a GitHub issue if problems persist

---

Quick Reference

Memory Modes:

• low_memory: 6GB VRAM, slower but stable
• balanced: 8-12GB VRAM, good speed+quality
• high_quality: 16GB+ VRAM, best quality

Recommended Settings by GPU:

GPU	Memory Mode	FP16	Max Resolution
RTX 2060 (6GB)	low_memory	True	432x768
RTX 3060 (8GB)	balanced	True	512x768
RTX 3070 (8GB)	balanced	True	720x1024
RTX 4080 (16GB)	high_quality	False	1080x1920

---

You're all set! Happy colorizing! 🎨