LayerUtility: CropByMask V4

BoundedImageCropWithMask gives you a tight rectangle around a mask. CropByMask V4 is the smarter, more opinionated version: it gives you three different ideas of what "around the mask" should mean, lets you add reserve margins and round the crop to a friendly multiple, and - the real feature - it can reuse one computed crop box across an entire video, so every frame gets cropped identically instead of the box jittering as the mask shifts.

This is a port of the LayerStyle/RunningHub RunningHub/LayerUtility node, and V4's two additions over earlier versions are exactly those: reuse_crop_box (compute once, cache, reuse) and a device choice for GPU-accelerated mask blur. It pairs naturally with the pack's RestoreCropBox V2, which pastes the result back using the same crop_box.

The three detect modes

• mask_area - the plain bounding box of the mask's non-zero region. What you'd expect.
• min_bounding_rect - the minimum bounding rectangle: the smallest rotated-into-place rectangle that contains the mask. If your mask is a diagonal person or a tilted object, this hugs it much tighter than an axis-aligned box.
• max_inscribed_rect - the largest rectangle that fits entirely inside the mask. Good for cropping to a clean interior region when the mask is an irregular blob and you want no background leaking in.

Inputs and outputs

Required: detect (the mode above), top/bottom/left/right_reserve (margins to add, default 20, negatives allowed), round_to_multiple (8–512 or None), reuse_crop_box, device. Optional: image, mask_image, and crop_box (if you already have a box, feed it and it's used directly).

Outputs: cropped_image, cropped_mask, crop_box (the BOX for restore/other nodes), and box_preview - an image with the detected region drawn in red and the final crop box in green. That preview output is worth wiring to a Preview node: it's the fastest way to see what the node thinks it's cropping before you trust it on a long batch.

How it works

It blurs the mask (Gaussian, radius 20) to smooth noise, runs the chosen detection, adds the reserve margins, rounds the width/height up to your multiple (centering the extra), and caches the result when reuse_crop_box is on. The first frame computes; every subsequent frame reuses the cached box. It also validates its inputs with clear errors - no mask, no box, no cache, and it tells you.

Installing it

It ships in ComfyUI_Swwan:

cd ComfyUI/custom_nodes
git clone https://github.com/aining2022/ComfyUI_Swwan
pip install -r ComfyUI_Swwan/requirements.txt

Or via ComfyUI Manager, search "ComfyUI_Swwan".

Where people get burned

reuse_crop_box is a class-level cache, so if you've got two instances of this node in one workflow, they share the cached box - and a change in one can silently affect the other. If the crop suddenly jumps for no reason, that's the culprit. And the cache persists for the process, so after you tweak a mask, clear the cache (reload the workflow or restart) before trusting the reuse path. The round_to_multiple centering can also push the box past the canvas edge on small images; the node clamps, but tight crops can end up off-center - check the box_preview before you commit.