Picker (SEGS)

Picker (SEGS) lets you choose, by hand, which of the detected regions actually go forward. A detector finds everything - five faces, a couple of false positives in the foliage, a hand you don't care about - and normally the detailer refines all of it. Picker sits in between and says: only these ones. You point at the SEGS you want and drop the rest.

This solves a real, specific annoyance in Impact Pack's detect-crop-refine pipeline. Detectors are indiscriminate; they'll flag a pattern in the background as a face, or catch three people when you only want to fix the one in front. Refining the junk wastes time and, worse, can "fix" things that weren't broken - turning leaves into a smeared pseudo-face. Picker gives you the veto. It also handles the flip side: when you crank a SEGSDetailer's batch_size to generate several candidate versions of a region, Picker is how you select the good one from the batch.

How it works

The node has a picks input - a multiline STRING where you enter which SEGS to keep - plus the segs themselves. In practice you run the graph once, and Picker opens a dialog showing the detected candidates so you can click the ones you want; your selection becomes the picks. If you select nothing, it outputs an empty SEGS (nothing gets detailed), which is a safety behaviour, not a bug - no selection means no work.

There's also an optional fallback_image_opt. SEGS carry mask information but not always image data; if you're previewing before a detailer has filled them in, connecting the original image here lets Picker show you the regions against the real picture instead of blanks.

The output is a filtered SEGS - the same structure, minus everything you didn't pick - which flows straight into a SEGSDetailer, an upscaler, or a paste-back node.

Why it beats just detailing everything

Two reasons. Control: you decide what matters instead of trusting the detector's threshold to get it exactly right. And curation: with batching, you generate options and keep the winner, which is a much nicer loop than re-rolling the whole image hoping the one face you cared about comes out right.

It's the manual counterweight to automation. The whole appeal of FaceDetailer and friends is that they run hands-off - but hands-off means the machine's judgment, and sometimes the machine picks wrong. Picker is where you put a human back in the loop for the one decision that actually needs taste.

Worth knowing this is one place ComfyUI genuinely one-ups the A1111/ADetailer way of doing things. ADetailer inpaints every detection on a shared seed, which is exactly how a crowd scene turns into a room of identical clones - you don't get to say "just that one." Picker (plus batching) hands you that selectivity directly.

Common issues

The classic confusion is the empty output: run it, don't select anything, and downstream nodes get an empty SEGS and appear to "do nothing." That's the node working as designed - make a pick. The other thing to remember is that Picker is interactive, so it wants you present to choose; in a fully automated, unattended run it's the wrong tool, because there's nobody to click. For lights-out selection by rule (biggest region, leftmost, highest confidence) use the SEGS filter nodes instead. Picker is for when you, a human, want to look and choose.

Installing it

Picker (SEGS) ships with ComfyUI Impact Pack. Install the pack via ComfyUI-Manager (search ComfyUI Impact Pack, Install, restart), or manually: cd ComfyUI/custom_nodes && git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack, then install the pack's requirements.txt into ComfyUI's Python environment (portable: ..\..\..\python_embeded\python.exe -m pip install -r requirements.txt), and restart. Auto-install ended in v7.6, so a manual clone needs the requirements step or the pack won't load; on first run it also fetches a SAM model into ComfyUI/models/sams. To produce the SEGS this node filters, you'll want a detector too - the YOLO provider moved to the separate ComfyUI-Impact-Subpack as of v8.0. Impact Pack is by ltdrdata, who also maintains ComfyUI-Manager, so it's well-supported.