✂️ Magic Mask Edit 遮罩裁切、缩放与回贴

If you've read anything about inpainting lately you've met the crop-and-stitch trick: cut the masked region out, let a model redraw just that patch at full resolution, paste it back. It became the 2026 default because the unmasked pixels never go through VAE encode/decode and come back bit-identical - a whole-frame edit model can't promise that.

MagicMaskEdit is that trick inside shigjfg's Magic Assistant pack - a Chinese-authored 12-in-1 utility suite. Same pattern as lquesada's Inpaint Crop and Stitch, different knobs.

One node, two jobs

Confusing on first open: you place two MagicMaskEdit nodes. Before your edit model, mode: crop. After it, mode: merge. Same class, two roles - these used to be two separate nodes, and both remain, deprecated, for old workflows.

In crop mode the node takes every pixel above mask_threshold, boxes it, adds padding pixels of context, shapes the box per crop_shape, then resizes it to fit inside target_width × target_height and rounds the canvas up to a multiple of size_multiple, filling the slack by replicating edge pixels. If your subject is jammed against the frame edge, the window slides inward instead of padding the source with black.

The clever part is what it saves. With the cropped image and mask, crop emits a crop_info blob carrying the original image tensor, the original soft mask, the crop box and the content box. So merge needs no original wired back in: one wire holds the geometry and the pixels. Merge resizes the edited patch to the crop box and composites with torch.lerp, weighted by that saved soft mask - everything else untouched. Note the deliberate asymmetry: mask_mode: binary hands the model a hard-edged, speck-preserving mask, while merge still blends with the soft original.

The settings that actually matter

Most of the nineteen inputs are situational. These are the ones you'll set:

• padding (64) - context halo in original pixels. Too little and the model has no surroundings to blend into; too much defeats the point.
• resize_mode + target_width/target_height (1024×1024) - how big the patch gets. In target mode aspect is preserved, the canvas is the full target, and the slack is edge-padded.
• feather (16) - inward feathering per region, in original pixels; 0 gives a hard edge. Never bleeds outside the mask.
• input_layout (canvas) - the one that bites. Use content only if the patch coming back has had its replicate padding stripped.
• empty_mask_mode (preserve) - no mask means keep the original, or edit the whole frame and restore the size.

Outputs are image, mask, crop_info and mask_empty. Crop's image and mask feed your edit model, its crop_info feeds merge, merge's image goes on to preview or save. mask_empty is the underrated one - gate the expensive branch on it.

The node face hides widgets that don't apply to the current mode; eight presets and the rest of the options live behind the ⚙ Settings & presets button.

Install

ComfyUI Manager → search Magic Assistant, or:

cd ComfyUI/custom_nodes
git clone https://github.com/shigjfg/ComfyUI-Magic-Assistant.git
cd ComfyUI-Magic-Assistant && pip install -r requirements.txt

The declared dependencies are just openai and aiohttp, and MagicMaskEdit needs no model files - the Klein and SDNQ loaders are the heavy ones. Keep ComfyUI on 0.29.0 or newer; the pack moved onto the new CoreModelPatcher APIs in v1.4.0. This node landed in v1.4.6; if it's missing, update the pack.

Where it breaks

Your edit came back identical. That's the mask-emptiness path: with empty_mask_mode: preserve, merge deliberately skips the paste when it found nothing above mask_threshold (0.001 by default - raise it and soft masks start failing the test). Check mask_empty, and read the status line the node prints on its face: crop box, output size, and whether it treated the mask as empty.

The paste is offset or zoomed. input_layout doesn't match reality: padding still in the image means canvas, padding gone means content.

Original image size differs from crop_info. You fed merge a different or resized original. Don't - crop already passed the original forward inside crop_info.

The edit model rejects the size. The 8-pixel canvas alignment suits SDXL and Flux; Qwen-Image-Edit-lineage models want multiples of 112, so set size_multiple to 112 and the target to 1008×1008. The canvas rounds up to a multiple of the alignment, so that lands cleanly.

Candidates won't merge. Merge takes an edited batch that matches the original, is a single image, or is an integer multiple of it - that's how you paste four variants of one inpaint in a pass, with batch_order choosing A B A B (repeat_batch) or A A B B (repeat_each).

No module named 'scipy'. The crop code imports scipy.ndimage for its feathering distance transform, and scipy isn't in the pack's requirements. It usually arrives with ComfyUI; lean installs don't. pip install scipy.