EricRollei-Eric_Image_Processing_Nodes/Docs/BM3D_Resolution_Scaling_Guide.md

BM3D Film Grain Denoising - Resolution Scaling Guide

Overview

The BM3D Film Grain Denoising node has been optimized for high-resolution images (1-20MP) with automatic parameter adjustment based on image resolution. This addresses the original issue where the denoising was too strong for large images.

Key Improvements

1. Resolution-Aware Sigma Scaling

The node now automatically adjusts the denoising strength based on image resolution:

Resolution	Scale Factor	Effective Sigma Reduction
≤ 2MP	1.0	No reduction (full strength)
2-4MP	0.9	10% reduction
4-8MP	0.75	25% reduction
> 8MP	0.6	40% reduction

2. Updated Grain Strength Presets

Base sigma values have been optimized for high-resolution processing:

Preset	Base Sigma	Best Use Case
ultra_light	5.0	High-resolution images (>8MP)
light	8.0	Subtle grain removal
medium	15.0	Balanced denoising (default)
heavy	25.0	Strong grain removal
custom	1.0-80.0	User-defined value

3. Adaptive Patch Size Scaling

The patch size automatically scales based on image dimensions:

• Small images (512x512): 3x3 patches
• Medium images (1024x1024): 7x7 patches
• Large images (2048x2048): 11x11 patches
• Very large images (4096x4096): 15x15 patches (capped)

4. Performance Optimizations

• Fast Mode: Automatically enabled for images > 4MP
• Conservative Scaling: Prevents overly aggressive processing on very large images
• Texture Preservation: Enhanced texture preservation factor (0.8 vs 0.7)

Usage Recommendations

For High-Resolution Images (>8MP)

• Use ultra_light or light grain strength
• Enable preserve_texture for natural film look
• Consider custom sigma values between 3-8 for very large images

For Medium Resolution Images (2-8MP)

• Use light to medium grain strength
• Standard settings work well with automatic scaling

For Low Resolution Images (<2MP)

• Use medium to heavy grain strength
• No automatic scaling applied (full strength)

Test Results

The scaling improvements were validated with test images:

Testing 512x512 (0.26MP)
  ultra_light: Effective Sigma: 4.0, Patch Size: 3x3
  light:       Effective Sigma: 6.4, Patch Size: 3x3
  medium:      Effective Sigma: 12.0, Patch Size: 3x3

Testing 1024x1024 (1.05MP)
  ultra_light: Effective Sigma: 4.0, Patch Size: 7x7
  light:       Effective Sigma: 6.4, Patch Size: 7x7
  medium:      Effective Sigma: 12.0, Patch Size: 7x7

Testing 2048x2048 (4.19MP)
  ultra_light: Effective Sigma: 3.6, Patch Size: 11x11 (0.9 scale factor)
  light:       Effective Sigma: 5.8, Patch Size: 11x11 (0.9 scale factor)
  medium:      Effective Sigma: 10.8, Patch Size: 11x11 (0.9 scale factor)

Technical Implementation

Sigma Calculation Formula

effective_sigma = base_sigma * resolution_scale_factor * texture_preservation_factor

Where:

• base_sigma: Preset-specific value (5.0, 8.0, 15.0, 25.0, or custom)
• resolution_scale_factor: 1.0, 0.9, 0.75, or 0.6 based on megapixels
• texture_preservation_factor: 0.8 if enabled, 1.0 if disabled

Patch Size Calculation

scale_factor = min(height, width) / 1000.0
# Conservative scaling for very high resolution
if scale_factor > 2.0:
    scale_factor = 1.0 + (scale_factor - 1.0) * 0.5
    
patch_size = max(3, min(15, int(7 * scale_factor)))

Migration Notes

From Previous Version

• Old light (sigma=15) → New medium (sigma=15) for same strength
• Old medium (sigma=25) → New heavy (sigma=25) for same strength
• New ultra_light (sigma=5) → Best for high-resolution images

Custom Sigma Recommendations

• 1.0-5.0: Very gentle (high-res images)
• 8.0-12.0: Light denoising
• 15.0-20.0: Medium denoising
• 25.0-35.0: Heavy denoising

Troubleshooting

If Denoising is Too Strong

• Try ultra_light preset
• Use custom sigma between 3-8
• Ensure preserve_texture is enabled

If Denoising is Too Weak

• Try heavy preset
• Use custom sigma between 20-35
• Disable preserve_texture for maximum removal

Performance Issues

• Fast mode automatically enables for >4MP images
• Consider downscaling extremely large images before processing
• Patch distance automatically increases for large images to improve performance

Future Enhancements

Potential improvements for future versions:

• Content-aware sigma adjustment based on noise level detection
• GPU acceleration for very large images
• Batch processing optimization
• Integration with other denoising methods for hybrid approaches