LC Character Detailer (BETA) 🩹

The problem it actually solves

A detail pass is the oldest trick in this hobby: the model has no latent budget for a face occupying 80 pixels of a 1024px frame, so you crop that face, re-render it at a proper resolution and paste it back. Impact Pack's FaceDetailer made that routine - right up until your image has two people in it, because a detailer treats every detected face identically. Same model, same prompt, same seed. Do that to a crowd and you get a room of clones.

LC Character Detailer is the blunt fix. SAM 3 finds every match and only one of them gets redrawn. You run one of these per character: branch the model, put that character's LoRA on the branch, point the node at their face. Two people, two nodes, no crossover.

How it works

No YOLO weights involved - the finding is promptable segmentation. You type simple words (face, hands) and SAM 3 returns a mask per instance; under the hood this node asks ComfyUI's built-in SAM 3 detector for up to 16 instances per word, so "hands" really does find both hands instead of the first one.

Then it measures every mask: centre of mass and area. That's what pick sorts on - left to right numbers them by horizontal centre, largest first by area - and number says which one this node cares about. Extra words like face, hair only come along when they belong to the picked match: at least 60% of the hair mask has to sit inside the face mask's box grown by its own size. That's how you get her hair rather than the hair of the person behind her.

From there it's the LC Smart Detailer engine: crop with padding context, scale to roughly inpaint_resolution squared pixels, masked redraw, paste back with a feathered edge and the seam colour matched. The preview numbers every match the way pick counts them, the redrawn one highlighted, so you can read the right number straight off the image.

The inputs worth setting

Defaults are sensible; these are the ones you'll touch:

• prompt - what to find, default Face. Comma-separated words, and face:1 caps it.
• pick + number - which match gets redrawn. in box is the interesting one: it takes elements from LC Bounding Boxes (a node in the companion LC123 pack), and the face stays with its character even when the pose changes.
• denoise - 0.3 is the author's starting point for faces. Low keeps the shape and fixes detail; high redraws from scratch and drifts the identity you're protecting.
• inpaint_positive (optional) - a short prompt for the crop, like close-up of a woman's face, natural skin. If your positive describes the whole scene, at denoise 0.3 and up the model will paint a tiny version of that scene into the face.
• cfg - defaults to 1 and refuses to go lower (below 1 the redraw is speckled). 1 is fine for Flux, Krea 2, Z-Image and even SDXL at detailer denoise. Wire negative only if your model uses one.

Outputs are image and mask. The mask is worth wiring somewhere: feed it to protect_mask on the pack's skin-detail nodes so they don't re-grain what the detailer just fixed.

Install

Via ComfyUI Manager, search ComfyUI_LC_MaskMaker_nodes, or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC_MaskMaker_nodes
# restart ComfyUI

The pack's install.py only tops up transformers and huggingface_hub when they're missing, never touches torch/numpy/Pillow, and bundles no weights. Two things you do need:

1. A ComfyUI with built-in SAM 3 support. The pack calls comfy_extras.nodes_sam3; an older build gets "This ComfyUI has no built-in SAM 3 support. Update ComfyUI." Update and restart.
2. The model. First use downloads sam3.1_multiplex_fp16.safetensors (1.7 GB) into models/sam3. Your own sam3.pt in models/sam3, models/detection or models/checkpoints is picked up instead.

Common issues

"Nothing redrawn: only 2 found, there is no number 3." It did what you told it, and the image passed through untouched. Change number, and check the preview - it counts by centre, not in the order SAM 3 returned things.

Wrong character redrawn. Look at the preview before touching denoise - the numbers drawn on the node are the numbers it uses.

The redraw is drastic. Drop denoise rather than blend - blending old and new pixels can leave you with two sets of eyelashes.

Fair warning: this pack is young and the nodes need a local install, so on a hosted ComfyUI platform you're waiting on the host to add it.