deepDeband (Banding Removal)

Removes banding artifacts and false contours using deep learning (ICIP 2022). Works on 8-bit or 16-bit images. Trained on 51,490 banded/pristine image pairs. The -w variant uses weighted patch fusion for better quality, while -f processes patches directly.

By subraoul·Created 9 months ago·Updated 9 months ago· 3

image

debanded_image

◄modeldeepDeband-w►

◄strength1.00►

◄tile_size256►

◄tile_overlap128►

Categorybitdepth_enhancement

Inputs (5)

Name	Type	Default	Description
image	IMAGE		Input image from previous ComfyUI nodes. Accepts both 8-bit (0-255 values scaled to 0-1) and 16-bit (0-65535 values scaled to 0-1) images. The node automatically detects and preserves the input bit depth.
model	COMBO	deepDeband-w	Model variant: deepDeband-w (weighted patch fusion, better quality, recommended) uses bilateral weighting to fuse overlapping patch predictions for smoother results \| deepDeband-f (full/direct) processes patches independently, faster but may have visible seams on large images
strength	FLOAT	1.000–1	Debanding strength (0.0-1.0). Controls blending between original and debanded images. 0.0 = original image (no debanding), 1.0 = full debanding effect. Use lower values (0.3-0.7) for subtle enhancement or when preserving intentional gradients.
tile_sizeopt	INT	256128–512	Tile size for processing (default: 256). The model was trained on 256x256 patches. Larger tiles may reduce seams but use more memory. Smaller tiles use less memory but may have more visible boundaries. Only matters for images larger than tile_size.
tile_overlapopt	INT	1280–128	Overlap between adjacent tiles in pixels (default: 128). The real deepDeband-w uses 50% overlap (128px for 256px tiles). Higher values reduce visible seams at tile boundaries but increase processing time. Recommended: 128 pixels for best quality matching original deepDeband-w. Set to 64-96 for faster processing (may show slight seams).

Outputs (1)

Name	Type	Description
debanded_image	IMAGE	—