Shima SEGSampler

Impact Pack's DetailerForEach is the canonical detect-refine loop: for every detected region, crop it, run a fresh sampling pass at proper resolution, paste it back. It's powerful and fiddly - two dozen inputs, separate positive/negative conditioning, and most people end up with a wall of nodes to feed it. Shima SEGSampler is this pack's answer: a "mega node" that wraps the whole thing so you hand it an image, SEGS, and one short prompt string, and it does the rest.

The pack's framing is exactly right - it abstracts the DetailerForEach complexity down to a single prompt input. You type what you want in the region into positive_add ("detailed features" by default, with negative_add set to "blurry, generic") and the node CLIP-encodes those strings internally, feeds them into DetailerForEach.do_detail, and returns the stitched result.

The inputs that matter

• image and segs - the base image and what to refine. Note: single image only; a batch raises an error rather than quietly doing the wrong thing.
• positive_add / negative_add - the strings appended to the region's re-render. This is the whole ergonomic win: you're writing a description ("detailed features"), not wiring conditioning.
• seed, steps, cfg, sampler_name, scheduler, denoise - the standard sampling stack. denoise defaults to 0.5, which is the sane start for a detail pass.
• model, clip, vae - connect individually, or wire a modelcitizen.bndl bundle and the node extracts all three. A bundle is the tidier path in this ecosystem.
• inpaint_model - optional dedicated inpainting model; if provided it's used for the detail passes instead of the main model.
• guide_size (default 512), max_size (1024), feather (5) - the crop/region sizing knobs inherited from the detailer. These control how big the re-rendered region is and how softly it blends back.
• The upscale* block - upscale, upscale_by, upscale_method, upscale_denoise, upscale_steps, upscale_cfg - an optional whole-image upscale pass after detailing.

Output is a single image. That's it. Detail + upscale + stitch, one wire out.

Installing it

Part of KDB-USJP/shima_wf:

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf

Restart ComfyUI. This node imports impact.impact_pack and calls DetailerForEach.do_detail directly, so ComfyUI-Impact-Pack is mandatory (the pack installer clones it). The optional upscale pass uses a standard UPSCALE_MODEL loader from core ComfyUI; no hidden downloads.

The honest trade-offs

If you don't supply a modelcitizen.bndl, you must wire model, clip, and vae individually or it raises a clear error - there's no silent fallback. And "mega node" cuts both ways: the two dozen DetailerForEach knobs are still there, just collapsed, and when a detail pass looks wrong you'll find yourself expanding back out to debug which crop setting did it. It's a great everyday wrapper, but it's a wrapper - the underlying Impact Pack machinery is what's actually doing the work.