AutoCropBorders

AutoCropBorders trims uniform borders from photos—think scanner edges, letterbox bars, or projector frames—before the image moves deeper into your workflow.

Inputs

image – The photo you want to clean; connect any RGB image tensor.
fuzz_mode – Choose how tolerant the detector should be:
- adaptive widens the tolerance automatically based on border noise.
- percent sticks to the value from fuzz_percent.
fuzz_percent – Maximum colour variation allowed when percent mode is active. Lower values assume a perfectly uniform border.
adaptive_k – Multiplier used in adaptive mode; higher values accept noisier borders.
edge_margin_px – Width of the strip sampled on each side to learn the border colour. Increase for thicker frames.
pad_px – Expand the final crop box by this many pixels if you want a little breathing room.
use_luma_only – When True, the decision runs on brightness only instead of full RGB. Handy for black-and-white content.
max_growth_iter – Safety cap on how many expansion steps the border mask can take.
return_border_mask – Toggle to output the detected border mask alongside the cropped image.
use_gpu – When True, temporary tensors jump to CUDA. Disable on tight VRAM budgets.

Outputs

image – The cropped photo with even width and height where possible.
left, top, width, height – Crop rectangle relative to the original image.
border_mask – Binary mask of removed borders (zeroed out when the mask output is disabled).

Where It Fits

Use AutoCropBorders on scanned headshots or production stills before resizing or running diffusion. It quickly removes black bars from footage exports so downstream nodes don’t waste cycles on empty pixels.

Tuning Tips

Switch to percent mode with a low fuzz_percent when dealing with perfectly solid borders (e.g., exact RGB black).
Bump edge_margin_px for thicker frames so the sampler grabs enough context.
Disable use_gpu if you are processing long batches on a shared card; the CPU path is slower but reliable.

Troubleshooting

Border remains – Increase fuzz_percent (in percent mode) or adaptive_k (in adaptive mode) so minor noise doesn’t block detection.
Image cropped too tightly – Raise pad_px or reduce the fuzz tolerance so fewer pixels are marked as border.
Unexpected slivers left behind – Check that max_growth_iter isn’t too low for high-resolution images; raising it lets the mask reach the true content edge.

Screenshot: docs/screenshots/auto_crop_borders.png

2.6 KiB Raw Blame History Unescape Escape