Prompt Chain Regional Detailer

The detect-crop-resample-composite loop behind ADetailer and Impact Pack's FaceDetailer is the standard fix for faces that come out mangled because they only occupy a few dozen pixels of the frame. What neither of those tools solves on its own is a scene with several faces that each need a different prompt - a plain face detailer redetails every detected face with the same shared text. Regional Detailer is PromptChain's answer to that specific gap: it pairs each detected (or posed) face with its own character's prompt from the regions data, then runs the same detect-crop-refine loop per face.

How it works

Two ways to find faces, chosen with face_source. detect (yolo) runs a YOLO detector over the rendered image - the same family of model ADetailer and Impact Pack use - fast, but it can miss a tilted or full-profile head. mannequin heads skips detection entirely and uses the exact head positions exported from the 3D Poser's pose data, which never misses a face the Poser knows about, frontal or in profile, as long as it's facing generally toward the camera. Once a face region is found, it's cropped with padding, upscaled to a working resolution, resampled with its own conditioning, and feathered back into the original at full size - the same mechanism as PromptChain_MaskedDetail, just with automatic detection and per-character prompting layered on top.

The inputs and outputs that matter

The tooltip on model is worth reading literally: it says a base model is preferred over an already-couple-patched one, because each crop is a single region, so plain regional conditioning is cleaner here than reusing an attention-patched model from an Attention Couple pass upstream. clip, vae, and image are the usual sampling trio; regions is Prompt Chain's 4th output, carrying each character's prompt.

Fields worth setting deliberately: denoise (default 0.4) controls how much of each face is actually re-rendered - lower keeps it closer to what was already there. face_source picks the detection mode above, and if you use mannequin heads, wire the optional pose input or the node has nothing to place heads with. max_face_angle (default 110°) only matters in mannequin mode - it culls heads turned more than that many degrees from camera, and the default is generous enough to keep frontal, three-quarter, and full-profile faces, only dropping ones facing fully away. bbox_threshold and bbox_dilation only matter in detect mode - confidence cutoff and how far to grow a detected box. mask_dilation grows the figure's silhouette (not the face box) before it bounds the re-render, so hair or headgear extending past the mannequin's edge doesn't get clipped. drop_size skips any detected face below that many pixels - useful for ignoring background extras. The optional negative input takes a shared negative if you have one; leave it unwired and it defaults to whatever negative is already in the regions JSON.

Single IMAGE output - the composited result, ready for your next step, whether that's an upscale or straight to output.

How to install it

Manager: search ComfyUI-PromptChain. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/mobcat40/ComfyUI-PromptChain.git

Face detailing is one of the pack's self-installing optional features - expect a first-use prompt to install its own detector and any missing model files, with progress shown live, rather than a separate Manager step.

Common issues & troubleshooting

A profile or tilted face gets skipped entirely. That's detect (yolo) doing what YOLO detectors always do with off-angle faces - miss them. If you posed the scene in the 3D Poser, switch to face_source: mannequin heads and wire pose; it works from known head positions instead of hoping the detector finds them.

Faces across a crowd start looking similar. This node takes one seed for the whole pass, same as most detailer implementations - when several faces are detailed in one run off a single seed, expect some convergence in how they come out, the way any single-seed multi-face detailer behaves. Rerun with a different seed on the outlier if one face needs to stand apart from the rest.

A face's prompt doesn't match the character. Confirm regions is actually wired from Prompt Chain, and if you're on mannequin-heads mode, that pose is wired too - name binding runs through the same $name{} mechanism as the rest of the pack.