Crop By Mask (UTK)

The detect-crop-refine pattern needs one thing done right: a crop box that actually hugs the subject. Crop By Mask (UTK) computes that box from a mask instead of relying on you guessing coordinates, crops the image and the mask together, and - usefully - gives you the crop box and a preview so you can see what it decided. It's the crop engine behind the pack's inpaint/restore workflow, and it's a faithful, if simpler, take on the mask-crop utilities that Impact Pack made famous.

How it works

Feed it an image and a mask_for_crop. The detect mode picks how the crop box is found:

• min_bounding_rect (the default behavior) - the tightest rectangle that contains the masked region. What you want 95% of the time.
• max_inscribed_rect - the largest rectangle that fits inside the mask. For ring- or donut-shaped masks where a bounding box would include the hole.
• mask_area - effectively the same bounding-rect math in this implementation; a compat option.

invert_mask flips which region counts, top_reserve / bottom_reserve / left_reserve / right_reserve (each default 20) add margin on each side so the subject isn't jammed against the crop edge - bump these up when faces get chopped.

Outputs

Five, and the last two are what make this usable in a loop:

• croped_image and croped_mask - the aligned crop pair, ready for a focused inpainting or upscale pass.
• crop_box - the coordinates, for pasting the result back into the original frame.
• box_preview - the image with the crop rectangle drawn on it, so you can eyeball the box before committing.
• remaining_area - the mask of what's outside the crop, useful for restoring the background later.

The pack even pairs this with a restore-crop-box node, which is the classic round-trip: crop → fix → paste back.

Installing it

Standard pack install - ComfyUI Manager → search ComfyUI-UniversalToolkit, or:

cd ComfyUI/custom_nodes
git clone https://github.com/whmc76/ComfyUI-UniversalToolkit
cd ComfyUI-UniversalToolkit && pip install -r requirements.txt

Then restart. PIL/OpenCV/numpy - all in the pack's requirements.

Gotchas

• The reserve values are pixels, not percents - 20 means 20 pixels on each side. On a 512px crop that's meaningful; on a 2048px one it's barely a hair. Scale them to your image size.
• An empty mask falls back to the full image (the bounding-rect code returns the whole frame when nothing matches). Forgiving, but it means a failed detection silently produces a full-frame crop - check box_preview, don't trust that a crop happened.
• max_inscribed_rect is a simplified implementation - in this pack it effectively returns the bounding rect, not a true maximal inscribed rectangle. If you genuinely need that algorithm, Impact Pack's crop-by-mask is the more complete tool; use this for the common bounding-box case.

For the standard "mask → tight crop → refine → paste back" workflow, this is a solid, self-contained version. It won't replace Impact Pack for the hard cases, but it's enough to build the loop without installing a heavy pack - and the box_preview output alone is worth it for debugging crops.