Node packVerifiedExternal

face-upscaling-and-seamless-embedding

FUSE KSampler is a face-aware KSampler node for ComfyUI that provides seamless face upscaling and embedding capabilities using YOLO face detection and SAM (Segment Anything Model) for precise masking.

Sourcegithub.com/WASasquatch/face-upscaling-and-seamless-embedding

FamilyYOLOMaintainer@WASasquatchUnratedgrade

Package profile

Status: Verified
Publisher: @WASasquatch
Source: GitHub

README

Face Upscaling and Seamless Embedding (FUSE)

FUSE is an All-in-One Face Fix KSampler for ComfyUI that provides seamless face upscaling and embedding capabilities using YOLO face detection and SAM (Segment Anything Model) for precise masking.

Before and After Examples

Installation

Navigate to your ComfyUI's custom nodes

cd custom_nodes

Clone FUSE nodes.

git clone https://github.com/WASasquatch/face-upscaling-and-seamless-embedding

Navigate to the FUSE directory

cd face-upscaling-and-seamless-embedding

Install dependencies to your Python Environment used by ComfyUI

pip install -r requirements.txt

Models

YOLO Face Models

You'll need to download some YOLO Face models. You can find some from:

Models should be placed in the models/yolo-face/ directory.

YOLO Models

You'll need to download some YOLO models. You can find some from:

Models should be placed in the models/yolo/ directory.

SAM Models

You'll also need to download some SAM models. You can find official models from:

https://github.com/facebookresearch/segment-anything?tab=readme-ov-file#model-checkpoints

Models should be placed in the models/sams/ directory.

Nodes

FUSEKSampler

Advanced face-aware sampling node that combines YOLO face detection, SAM segmentation, and seamless blending. Features include:

Intelligent face detection and cropping with adjustable padding and size
Optional square crop enforcement for consistent face processing
Multiple blending modes (box/radial) with configurable strength
SAM-based precise face masking
Color preservation options using various transfer methods (LAB, PDF, mean-std)
Internal caching system for faster workflow iteration
Support for processing specific faces using face index

Parameter	Description	Type	Default	Value Range
model	Base model for sampling	MODEL	Required	N/A
vae	VAE model for the sampler	VAE	Required	N/A
images	Input image batch for face detection and upscaling	IMAGE	Required	N/A
positive	Positive conditioning for the sampler	CONDITIONING	Required	N/A
negative	Negative conditioning for the sampler	CONDITIONING	Required	N/A
use_cache	Use internal caching to speed up workflow iteration	BOOLEAN	True	True/False
seed	Seed for deterministic results	INT	0	≥ 0
steps	Number of steps for the sampler	INT	20	≥ 1
cfg	Classifier-Free Guidance scale	FLOAT	8.0	Any float value
sampler_name	Name of the sampler	STRING	(from KSampler)	Available KSampler options
scheduler	Scheduler type	STRING	(from KSampler)	Available KSampler schedulers
denoise	Denoising strength (lower values adhere more to input face)	FLOAT	0.5	0.0 - 1.0
yolo_detector	YOLO model for face detection	STRING	Required	Available YOLO face models
yolo_confidence	YOLO confidence threshold (higher values detect only more confident faces)	FLOAT	0.25	0.0 - 1.0
yolo_nms_iou	YOLO NMS IoU threshold (lower values remove more overlapping detections)	FLOAT	0.5	0.0 - 1.0
sam_segmenter	SAM model for face segmentation	STRING	Required	Available SAM models
sam_model_type	SAM model type	STRING	"vit_b"	"vit_b", "vit_l", "vit_h"
face_id	Index of the face to process (0 is first face, -1 is all faces)	INT	0	≥ -1
face_order	Order to process detected faces	STRING	"linear"	"linear", "linear_reve