LC Smart Detailer (pipe) 🩹

This is LC Smart Detailer with the settings supplied by a pipe instead of by hands. Same detection, same per-region crops, same blend. The difference is what's on the node: no model, no vae, no steps, no cfg, no sampler, no scheduler, no seed, no positive wire - those arrive through pipe, and the node passes the pipe straight back out so you can chain the next thing off it.

Whether that's better depends entirely on whether you already live in the LC pipe world. If you're running the companion ComfyUI_LC123_nodes pack and your graph is built around pipes, this is the version that stops you re-wiring seven connections into every detail pass. If you're not, use the non-pipe node - reading the source, this one pulls its values from pipe keys, so it's only useful if something upstream is actually writing them.

What comes from the pipe and what doesn't

The pipe supplies model, vae, seed, the step count, cfg, sampler_name, scheduler, and conditioning (positive and negative). The step count it reads is the detailer-oriented step key, falling back to whatever is in the pipe - which is the small tell that the author designed the pipes for detailers in the first place. The cfg the pipe gives is raised to a minimum of 1.0 on the way in, same as on the non-pipe node: below 1 the redraw comes out speckled.

What stays on the node: prompt (what to find - hands, face, comma separated, :N to cap a word like face:1), sam3_model, threshold, and the shape of the pass - grow, feather, blend, padding, inpaint_resolution, denoise. Those are per-use decisions, so the author kept them local, which is the right call.

The optional inputs are the escape hatches. Wire model to use something else for the redraw - a model with LoRAs on it, or a different checkpoint entirely - and vae to swap the VAE. Wire positive and inpaint_positive if the pipe has no conditioning or if you want the crops to get a different prompt from the one that made the image; inpaint_positive is the one to reach for whenever positive describes the whole scene, because at denoise 0.3 and up that scene gets drawn into every crop. Wire negative for SDXL and other models that use one; leave it empty for cfg-1 models.

Outputs are pipe (continuation), image, and mask.

How the pass itself works, briefly

It's the same pipeline as its sibling: SAM 3 finds each thing you named - up to 16 objects per word, so "hands" gets both hands - masks are grown and merged when crops overlap, each crop is enlarged to about inpaint_resolution squared by area, redrawn as a masked inpaint with the seam ring, colour-matched to the original, and pasted back with a blend-weighted feather. denoise 0.3 is the detail-fix setting; 0.6 and up starts redrawing from scratch. Nothing found means the image passes through untouched and the node tells you. The non-pipe node's page goes into the tuning in more depth, and it all transfers - only the wiring differs.

Installing it

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

Or ComfyUI Manager, searching "LC MaskMaker" (ComfyUI_LC_MaskMaker_nodes, v0.17.1, MIT). The installer only checks for transformers and huggingface_hub. You also want the companion ComfyUI_LC123_nodes pack - that's where the LC pipe producers live, and without them there's nothing to plug into pipe. SAM 3 comes from ComfyUI itself: you need a build with built-in SAM 3 support, and the SAM 3.1 download (1.7GB, models/sam3) or your own sam3.pt / sam3.safetensors in that folder.

Where people get burned

"The pipe has no model, vae, seed, steps, cfg, sampler_name, scheduler." That's the node's own error, and it names the missing pieces exactly. Something upstream isn't filling the pipe - add it with the pipe creator/edit nodes from LC123 and run again. It's the single most common failure here, and it's not a bug.

Wiring a LoRA-laden model and expecting the pipe's model too. You get the wired one. That's the point of the override, but it's easy to wire a model "just to test" and then wonder why the redraw changed character.

Crowds. Every detected region is redrawn on the same seed, so a group of faces can converge on one person. Detail crowds by hand, or run the regions separately.

Over-processing. Detailing an already-sharp face spends render time to change an image without improving it. Fix at native resolution before upscaling, and skip the pass when the region wasn't broken.