FaceDetailer (pipe)

FaceDetailer (pipe) is FaceDetailer for when one pass isn't enough. It does the same thing - detect faces, crop them, resample at proper resolution, paste them back - but instead of accepting a model, VAE, prompt, and detector as separate wires, it takes them all bundled into a single DETAILER_PIPE, and it outputs that same pipe again so the next FaceDetailer can pick up where this one left off. That output-is-also-an-input design is the entire reason the node exists.

The problem it solves is the 2-pass detail restore. A severely mangled face sometimes needs two goes: a first pass at modest settings to recover the rough shape, then a second at higher settings to put the detail back. You can wire two full FaceDetailers in a row, but each one needs the model, clip, vae, positive, negative, and detector re-connected - a mess of a dozen duplicate wires. The pipe collapses all of that into one cable. Chain three FaceDetailer (pipe) nodes and you've re-plugged one wire between them, not twelve.

How it works

A DETAILER_PIPE is a bundle: model, CLIP, VAE, positive and negative conditioning, the BBOX detector, an optional SAM model, and wildcard settings, all packed together. You build one with ToDetailerPipe (or BasicPipe -> DetailerPipe), feed it in here, and the node unpacks it to run the same detect-crop-refine loop FaceDetailer runs. Then it re-emits the pipe on its output so the next stage inherits the identical setup. Everything past the pipe is standard sampling - the same masked, full-resolution regional resample that underpins every detailer in the pack, and the same idea as ADetailer over in A1111.

The inputs and outputs that matter

• image (required) - the picture to fix.
• detailer_pipe (required) - the bundle of model/clip/vae/conditioning/detector. Build it once with ToDetailerPipe upstream.
• guide_size (default 512) - the size each detected face is scaled to before resampling. Where the detail comes from.
• denoise (default 0.5) - how hard the pass rewrites the face. The main dial. In a 2-pass setup, keep the first pass low (rough recovery) and let the second do the detail; 0.3–0.4 is the safe band for staying on-model.
• feather (default 5) - softens the paste edge.
• bbox_threshold (default 0.5), bbox_dilation, bbox_crop_factor - detection tuning: confidence cutoff, mask growth, and how much context each crop keeps.
• cycle (default 1) - refine passes per face within this node.

The sam_* inputs only do anything when the pipe carries a SAM model, in which case they steer how SAM tightens each mask to the silhouette. Outputs: image (the fixed frame), detailer_pipe (the same bundle, for the next stage - this is the one that makes chaining clean), plus mask (a visualization of what got detailed), cropped_refined and cropped_enhanced_alpha (per-face crops for inspection), and cnet_images.

How to install it

Install the pack via ComfyUI Manager (search ComfyUI Impact Pack, install, restart), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack comfyui-impact-pack
cd comfyui-impact-pack
python -m pip install -r requirements.txt   # ComfyUI's python; python_embeded on portable

then restart. The face detector inside the pipe comes from the Impact Subpack (UltralyticsDetectorProvider), which you install separately since v8.0. The bundled SAM model auto-downloads to ComfyUI/models/sams on first install. Linux: libgl1-mesa-glx and libglib2.0-0 for OpenCV.

Common issues & troubleshooting

Just starting out? Use plain FaceDetailer first. The pipe version earns its keep only when you're chaining passes. For a single face fix, the one-node FaceDetailer has fewer wires to get wrong; move to the pipe when you actually need 2-pass.

Both passes barely changed the face. If the face was already large and sharp, a detail pass mostly costs time and drifts the style - the most-upvoted reaction to a detailing demo was that the two versions looked identical. Detailers pay off on small faces; skip them on close-up portraits.

Second pass overcooked it. Lower the second stage's denoise. Stack two passes and their effect compounds - a 0.5 followed by another 0.5 is a lot of rewriting.

Nothing connects to detailer_pipe. You need to build the pipe upstream with ToDetailerPipe and give it a detector from the Subpack. Without a detector in the bundle, there's nothing to find faces with.