SEGS Preview

SEGS is the data format Impact Pack uses to hand regions around: a detector finds faces or hands, each one becomes a segment with its own cropped image, mask, bounding box, and crop region, and a detailer re-renders them one by one. The problem is you're flying blind while that happens. SEGS Preview [Eclipse] is the debug window - it unpacks that SEGS bundle and shows you each segment as its own cropped image, so you can see exactly what the detector grabbed and how tight the crops are before anything gets composited back.

Why you'd actually reach for this: you're running a Detailer (SEGS) or FaceDetailer workflow and a result comes back with a seam, a chunk of background that got re-rendered, or a face that was half-cropped. Those are detector problems, not sampler problems, and you can't diagnose them without looking at the segments themselves. This node makes that inspection a two-second glance instead of a guessing game.

How it works

Impact Pack's SEGS structure is basically (image_size, list_of_SEG), where each SEG bundles its own cropped_image, cropped_mask, crop_region, and bbox. This node iterates that list, crops each segment out, optionally turns the segment mask into an alpha channel, and saves the results as temporary previews. It's a port of the Impact Pack preview node, so if you've used that one, none of this is surprising - the Eclipse version just adds a couple of knobs.

The inputs that matter:

• segs - the SEGS data, straight from a detection node or a SEGS splitter.
• alpha_mode (on by default) - applies the segment mask as transparency. Off, and the crops come back as flat rectangles with their background baked in.
• min_alpha (default 0.2) - the floor for that alpha, so pixels outside the mask aren't fully transparent. Drop it toward 0 for a hard cutout.
• crop_padding (default 10) - padding around the tight crop. This one is worth touching: a face detector's crop is often right up against the chin, and bumping this to 30–50px gives the detailer room to work. Negative values shrink the crop inward.

The second output, tight_crop, is the useful quiet one: a single clean crop of the source image around all segments combined - no masking, no black bars. It's the "here's the whole region as one image" view, and it requires fallback_image_opt (the source image) to be wired in.

Outputs and where they go

IMAGE is a batch of per-segment crops. You can feed it onward for inspection, OCR, or any per-face processing, and the node itself is an output node, so the previews also just appear in your temp folder. tight_crop is a single image you can pipe into a second detailer or a save node.

Installing it

SEGS Preview ships in ComfyUI_Eclipse. Easiest is ComfyUI Manager → search "ComfyUI_Eclipse" → install, then restart. Or from your ComfyUI folder:

git clone https://github.com/r-vage/ComfyUI_Eclipse custom_nodes/ComfyUI_Eclipse

No model downloads - it's pure image processing. It does expect Impact Pack's SEGS type to be present, so have Impact Pack installed (you almost certainly do if you're producing SEGS).

Gotchas

Where people get burned: the tight_crop output silently needs fallback_image_opt - leave it unconnected and you get nothing where you expected a crop. And if a segment's cropped_image is missing (some detectors don't populate it), the node falls back to cropping the source, which again needs that input wired. One thing to know about this pack: Eclipse went through a v4.0.0 cleanup that removed legacy node IDs, so an old workflow referencing the older RvTools-era node names will show red nodes until you run the included migration tool. If you're building fresh, none of that touches you.