Inpaint Options

Person Detailer will happily detail every matched face with its default settings and you'll get good results on day one. The Inpaint Options node is for when you want to stop trusting those defaults. It's a config bundle that plugs into Person Detailer's inpaint_options input, and it hands you three things Person Detailer alone doesn't expose: a CFG scale, a negative prompt for the inpaint pass, and per-slot control over which mask type each person gets, how many inpaint rounds they get, and whether Detail Daemon runs for them.

Think of it as the difference between one-size-fits-all face detailing and the version where person one gets a face-skin-only pass while person two gets a full head re-render. Same node, wildly different control.

The three inputs that do the work

• cfg - Classifier-Free Guidance scale for the inpaint sampling. Default is 1.0, which means no guidance at all. That's not a mistake: it's tuned for distilled/turbo models where CFG over 1 fights the model. If you're on a non-distilled checkpoint and faces come out washed out or low-contrast, this is the dial - 1.5–3.0 fixes paleness, 4.0–7.0 is standard for regular SD models. Go past 8 and you're asking for artifacts.
• negative_prompt - text that's encoded and merged with whatever negative conditioning you feed Person Detailer. If you connect no negative at all, this text alone becomes the negative. It's the obvious home for "blurry, washed out, pale, low contrast, extra fingers, deformed hands" - the standard inpaint-artifact suppressors from the inpainting playbook.
• denoise_gradient / denoise_gradient_mode - a center-to-edge falloff for denoise. Set 0 and the whole mask gets uniform denoise (default). Crank to 0.5 and the middle of the mask gets full denoise while edges stay close to the original pixels, which kills the hard seam you get when an inpainted region doesn't blend with untouched surroundings. Modes: linear (smoothest, best default), radial (good for round face masks), smooth (follows the exact mask shape - the one to pick for body parts and accessories).

Per-slot configuration

This is where the node earns its keep. Each reference row (1–5 plus generic) gets four settings:

• reference_N_mask_type - face (skin only), head (face + hair, the default), body, hair, facial_skin, eyes, mouth, neck, accessories, or aux (body-part masks from a detector). Yes, you can detail one person's eyes and another person's whole body in the same pass.
• reference_N_rounds - number of inpaint passes for that slot. Each round re-encodes the previous result (latent cycling).
• reference_N_detail_daemon - toggles Detail Daemon sigma manipulation for that slot specifically.
• reference_N_ref_index - which reference (1–10) this row points at, so rows can be remapped if your Person Selector is set up with more references than slots.

For progressive refinement, the two global strings matter: denoise_progression and steps_progression. Pipe-separated values, one per round - 0.5|0.3 as denoise and 6|4 as steps gives you a strong first pass then a gentle refinement pass, which is the classic fix for faces that come out over-detailed or plastic-looking. If a slot has rounds of 1, progression is ignored and Person Detailer's global values win.

Installing and wiring

It ships with the FVMtools pack - install once, every node is there:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

No extra Python deps for this one (numpy's already in ComfyUI). Wire it up: Inpaint Options' single inpaint_options output → Person Detailer's inpaint_options input. That's the whole job - it doesn't touch the image itself.

The trap

The defaults hide a footgun: denoise_gradient at 0 with cfg at 1.0 means your inpaint pass runs with zero guidance and a hard-edged mask. That's fine on a turbo model at low denoise, but the moment someone runs this on Flux or SDXL and wonders why faces look washed out or seams show, these two sliders are the answer. Bump CFG to ~4, add a gradient of 0.3–0.5, and the same workflow suddenly looks like a different pack. Start there before you blame the model.