MASK to SEGS

Impact Pack's detailers and upscalers eat SEGS, not masks. So when you already have a mask - you painted it by hand, or you got it from CLIPSeg, or SAM, or a background remover - and you want the detailer to work on that exact area, you need a converter. MASK to SEGS is it. Hand it a mask, get back a SEGS you can feed straight into DetailerForEachPipe or SEGS Upscaler. It's the bridge that lets you skip the built-in detectors entirely and detail wherever you decide.

This is genuinely handy. It means the whole detect-crop-refine machine isn't limited to what a YOLO model can find. Anything you can mask, you can detail.

How it works

It reads the mask, finds the distinct blobs in it, and wraps each one in a SEG with a crop region sized around it. By default each separated region becomes its own SEG, so a mask with three painted spots produces three detections the detailer will handle independently. Flip combined and the whole mask collapses into one region instead.

The inputs and outputs that matter

• mask (MASK) - your source mask. Any mask node upstream works.
• combined (default off) - off means each separate blob becomes its own SEG; on means the entire mask is treated as a single region. Turn it on when you want one unified detail pass over everything you masked, rather than one pass per blob.
• crop_factor (default 3) - how much surrounding context each crop includes. Same idea as on the detectors: more context blends better but works a bigger area. 3 is a sane start.
• bbox_fill (default off) - fill the whole bounding box as the mask rather than following the painted shape.
• contour_fill (default off) - fill the interior of each contour, useful when your mask is an outline rather than a solid shape.
• drop_size (default 10) - ignore blobs smaller than this, so stray specks in a rough mask don't each become a detection.

The single output is SEGS - wire it into a Detailer, a filter, or a preview.

How to install it

ComfyUI Manager: search ComfyUI Impact Pack, Install, restart.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack
cd ComfyUI-Impact-Pack
pip install -r requirements.txt

Run that in ComfyUI's Python environment and restart. This node needs no detection model of its own - that's the whole point, you're supplying the "detection" as a mask. It does depend on the pack's normal libraries (OpenCV, scikit-image), which the requirements file installs.

Common issues & troubleshooting

One big mask became a dozen tiny SEGs. Your mask has speckle. Raise drop_size to ignore the little blobs, or turn on combined to force it all into one region.

The detail pass has a hard rectangular edge. You're probably running with bbox_fill on, which masks the full box instead of the painted shape. Turn it off so the region follows your actual mask outline, and give it some crop_factor for context to blend into.

A thin outline mask detailed nothing useful. If your mask is a hollow contour, turn on contour_fill so the enclosed area becomes solid before it's converted - otherwise there's barely any masked region for the detailer to work in.

Remember to composite. SEGS from a mask still run through the normal detailer, which crops, re-renders and pastes back. As always with this pattern, keep denoise moderate on the downstream detailer so the fixed region blends rather than reinventing itself.