SEGS Enhancer Multi (CRT)

SEGS Enhancer Multi (CRT) is the composable version of the CRT face detailer. Where Ultralytics Enhancer (CRT) runs its own YOLO detection, this node consumes a SEGS input - you bring the detections, it does the enhancing. That one swap is the whole story, and it's what makes this node more powerful than its sibling: SEGS can describe faces, hands, people, or anything else a detector or mask pipeline found, so this becomes a general "refine every detected region" node rather than a face-only tool.

SEGS is Impact Pack's structured segmentation format - a list of detected regions, each carrying a bounding box, a mask, and confidence. If you've built any Impact Pack graph, you've already got SEGS flowing around (from a BBOX or SAM detector provider, or a SEGM (Ultralytics) node). The KB treats SEGS as the composable backbone of ComfyUI detailing: Detailer (SEGS) is the go-to when you want detection and refinement controlled separately. This node is in that same family - it takes your SEGS, and for every entry it runs the same two-stage, noise-injection-enabled refine-and-composite loop as the face enhancer.

The execution per region is identical to its sibling: crop the region, upscale it to upscale_megapixel (default 1.5) so the refine pass has real budget, run the low-step sampler (defaults: lcm, cfg 1.0, steps 2 - tuned for distilled models), then feather the result back with enhancement_mix and color_match_strength over a masked blend controlled by post_mask_expand, post_mask_blur and post_mask_taper_borders. The full stage1/stage2 sigma and details_amount_* controls are here too, along with the noise-injection block (enable_noise_injection, injection_point, injection_strength, normalize_injected_noise) and edit_model_flux2klein for edit-model reference conditioning.

Inputs that matter: image, segs (from your detector), model, positive, and the sampler quartet - plus upscale_megapixel, resize_back_to_original, enhancement_mix, color_match_strength. Optional negative and vae again. The outputs are the same five as the face version: enhanced_image (use this), enhanced_face, cropped_face_before, and the enhanced_face_alpha / base_face_alpha layers for custom compositing.

The honest difference from the face-only node: garbage in, garbage out. Since detection is upstream, the regions you enhance are exactly what your SEGS source hands you - if your detector emits noise, this node will happily "enhance" noise. That's also its superpower: feed it SEGS for hands, or clothing seams, or product regions, and you've built a general region-detailer without writing a new node. The common trap is double-detecting - if you already ran a face-detailing pass, don't feed those same SEGS back in; the "second pass on an already-fine region" problem the KB flags applies here too.

Install is the pack's: ComfyUI Manager → search "CRT-Nodes", or clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart. You'll also want Impact Pack installed (for the SEGS generators) and, if you're doing your own detection with the pack's Ultralytics nodes, a model in models/ultralytics/segm or bbox. Given the December 2024 Ultralytics supply-chain incident documented in the KB, pin your ultralytics install to a post-incident version. If regions come back unchanged or the console complains about a missing SEGS producer, check that your detector's threshold is sane before blaming the enhancer.