DetailerFix

A face that takes up 80×80 pixels of a 1024×1024 image only gets 80×80 pixels' worth of the model's attention, and it comes out mangled - smeared eyes, mushy features. The fix everyone uses is the detect-crop-resample-paste loop: find the face, crop it out, run a fresh sampling pass on it at proper resolution, paste it back. easy detailerFix is Easy-Use's wrapper around exactly that loop (the same machinery as Impact Pack's FaceDetailer), packaged to work with Easy-Use's pipe.

If you've generated a portrait and the face looks off, this is the node that rescues it. It's not limited to faces either - with the right detector it does hands, eyes, anything a detection model can find. It's the standard "make the small stuff not look broken" pass, and it slots into an Easy-Use graph without you having to leave the pack for Impact Pack's node soup.

How it works

The detailing loop needs three things bundled together: the model/CLIP/VAE to sample with, and a detector to find the regions. In an Easy-Use graph you assemble those upstream - an easy preDetailerFix gathers the sampling settings, and easy ultralyticsDetectorPipe / easy samLoaderPipe supply the detector - and it all flows in as one pipe. easy detailerFix then, for each detected region: crops it, denoises it at full resolution, and composites it back into the image. Same idea as Impact Pack's FaceDetailer, because it's built on that lineage.

The inputs and outputs that matter

• pipe - the bundle carrying your model, conditioning, and detector setup (from easy preDetailerFix + a detector pipe). This is the required input, and getting it assembled is most of the work.
• image_output - Preview, Save, Hide, and combinations. How the result is surfaced.
• save_prefix - filename prefix if you're saving.
• model (optional) - override the model used for the detail pass.

Outputs: pipe (pass-through, chain another fix after it), image (the refined result), and cropped_refined / cropped_enhanced_alpha (the individual fixed regions, as lists - handy for inspecting what it actually changed).

How to install it

Part of ComfyUI-Easy-Use. ComfyUI Manager: search ComfyUI-Easy-Use, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat (Windows) or pip install -r ComfyUI-Easy-Use/requirements.txt, and restart. Important: the detection side depends on ComfyUI Impact Pack (and, for YOLO detection, its separately-installed Impact Subpack) plus a detector model. Install Impact Pack alongside Easy-Use, and grab a detection model like bbox/face_yolov8m.pt into your Ultralytics models folder - without a detector, the detail pass has nothing to find.

Common issues & troubleshooting

"No detector" / nothing gets refined. The most common failure: the detector pipe isn't set up, or the YOLO detection model isn't installed. This node can't find faces on its own - it needs easy ultralyticsDetectorPipe feeding the pipe and an actual .pt detection model on disk. Impact Pack (and its Subpack for the Ultralytics detector) has to be installed too.

The face changes but doesn't improve. If the face was already large and well-resolved, a second pass just rewrites it without making it better - the most common reaction to detailing an already-good face is "these look identical." Use this when the face is small in frame; skip it when it's already big.

Detail pass looks pasted-on. If the refined region doesn't blend, the denoise strength (set upstream in easy preDetailerFix) is probably too high, so the new face diverges from the body. Lower it so the pass enhances rather than reinvents.

Batch-size errors on old installs. The pack's changelog records fixes for detailerFix erroring when batch size was greater than 1. If you hit that, update Easy-Use.

A note on Impact Pack's dependency. The Ultralytics detector had a supply-chain incident once, which is why the risky detector now lives behind the opt-in Impact Subpack rather than being bundled. Install it deliberately, from the official source, and you're fine - just know that's why detection isn't included by default.