FUSE KSampler (Generic)

FUSE KSampler (Generic) is the same machine as the face version, with the face-specific naming scrubbed out so you can point it at anything a YOLO model can find. Faces are just the most famous thing that gets mangled at small sizes - hands are the other one, and objects, and anything you want to isolate, re-sample at proper resolution, and blend back in cleanly. Where FUSE KSampler reads face_size, face_id, and a face-trained detector from models/yolo-face, this one reads mask_size, mask_id, and pulls detectors from the general models/yolo folder.

If you're only fixing portraits, use the face version and don't think about this one. Reach for Generic when your problem isn't a face.

How it works

Mechanically identical to the face node: YOLO detects boxes on the input image, each box gets padded and optionally forced square, cropped, resized to mask_size (512 default, up to 1536), VAE-encoded, run through the normal ComfyUI sampler at denoise strength, decoded, and pasted back with feathered blending. The mask_id (-1 for all) and mask_order (linear, largest bbox, etc.) controls pick which detected region gets processed when there are several. Same caching by image hash, same images + masks outputs.

The difference is what you can hang off it:

• Hands. A hand detector (there are YOLO hand models in the same ecosystem) plus a model that's decent at hands, and you get a hands detail pass without buying into a hand-specific inpainting workflow. This is the most common reason people grab this node.
• General object isolation. Anything a generic YOLO model has classes for - cars, people, animals. You get the crop-and-fix loop plus class filtering via the yolo_optionals input.
• Per-class targeting. Combined with FUSE YOLO Optionals, set class_filter to the class you care about so the sampler ignores everything else in the frame.

Inputs that matter

• model / vae / images / positive / negative - same sampler wiring as always.
• yolo_detector - now sourced from models/yolo/ (general models), not models/yolo-face/. Different folder, different dropdown.
• sam_segmenter + sam_model_type - SAM is still there for contour-accurate masks instead of box feathering.
• mask_size - the sampling resolution for the crop. 512 is fast; 1024+ for a serious detail pass.
• denoise - lower sticks closer to the original region, which matters more here: you usually want to fix a hand, not invent a new one, so start around 0.4–0.5.

Optional mask_optionals and yolo_optionals inputs take the same DICTs as the face version.

Outputs: images (the fixed batch) and masks (the per-region blend masks).

Installing and the model folders

Same pack, same install - it's all in FUSE Face Enhancer by WASasquatch, installable from ComfyUI Manager or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/face-upscaling-and-seamless-embedding
cd face-upscaling-and-seamless-embedding
pip install -r requirements.txt

Restart, then put general YOLO models in ComfyUI/models/yolo/ (the face models live in models/yolo-face/; they don't cross over). SAM checkpoints go in models/sams/ regardless.

The traps

• Wrong folder. Putting a face model in models/yolo/ (or vice versa) and wondering why the dropdown is empty - the node only lists the folder it's wired to read.
• Detecting faces with a general model. A generic YOLO will happily call a face "person" and crop the whole torso. That's what class_filter and a proper model are for.
• Expecting the face version's defaults. The blending and color controls behave identically, but the thing you're fixing is no longer a face, so the "don't drift identity" advice transfers to "don't redraw the object." Keep denoise modest or you're doing generation, not repair.