nodes/SykkoAtHome-ComfyUI_FaceProcessor
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

README

Browse Source
This commit is contained in:
Rafał Tłuchowski
2025-01-26 13:41:11 +01:00
parent 1e248e60f7
commit db92f67158
3 changed files with 201 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
__init__.py
+1 -1
View File
@@ -19,7 +19,7 @@ NODE_CLASS_MAPPINGS = {
NODE_DISPLAY_NAME_MAPPINGS = {
"FaceFitAndRestore": "Face Fit or Restore",
"FaceWrapper": "Face Landmarks Detector"
"FaceWrapper": "Face Wrapper"
}
__version__ = "1.0.1"
nodes/face_wrapper.py
+42 -7
View File
@@ -48,6 +48,10 @@ class FaceWrapper:
image_np = self._convert_to_numpy(image)
height, width = image_np.shape[:2]
if mode == "Wrap":
return self._wrap_mode(image_np, None, width, height,
device, x_scale, y_transform, processor_settings)
# Detect facial landmarks
landmarks_df = self.face_detector.detect_landmarks(image_np)
if landmarks_df is None:
@@ -59,15 +63,10 @@ class FaceWrapper:
return self._debug_mode(image_np, landmarks_df, width, height,
show_detection, show_target, landmark_size,
show_labels, x_scale, y_transform, processor_settings)
elif mode == "Un-Wrap":
return self._unwrap_mode(image_np, landmarks_df, width, height,
device, x_scale, y_transform, processor_settings)
elif mode == "Wrap":
print("Wrap mode not implemented yet")
return (image, processor_settings or {})
def _debug_mode(self, image_np, landmarks_df, width, height, show_detection,
show_target, landmark_size, show_labels, x_scale, y_transform,
processor_settings):
@@ -133,8 +132,7 @@ class FaceWrapper:
warped_image = GPUDeformer.warp_face(
pil_image,
source_landmarks,
base_landmarks,
device='cuda'
base_landmarks
)
else:
warped_image = CPUDeformer.warp_face(
@@ -151,6 +149,43 @@ class FaceWrapper:
landmarks_data = self._prepare_landmarks_data(landmarks_df, base_landmarks)
return (output_image, self._update_settings(processor_settings, landmarks_data))
def _wrap_mode(self, image_np, landmarks_df, width, height, device,
x_scale, y_transform, processor_settings):
if not processor_settings or 'target_lm' not in processor_settings:
print("No landmarks found in processor settings")
output_image = torch.from_numpy(image_np.astype(np.float32) / 255.0).unsqueeze(0)
return (output_image, {})
# Source landmarks from stored base landmarks (target_lm in settings)
source_x = processor_settings['target_lm']['x']
source_y = processor_settings['target_lm']['y']
source_landmarks = np.column_stack((source_x, source_y))[:468]
# Target landmarks from detected face (detected_lm in settings)
detected_x = processor_settings['detected_lm']['x']
detected_y = processor_settings['detected_lm']['y']
target_landmarks = np.column_stack((detected_x, detected_y))[:468]
# Select warping method
pil_image = Image.fromarray(image_np)
if device == "CUDA" and torch.cuda.is_available():
warped_image = GPUDeformer.warp_face(
pil_image,
source_landmarks,
target_landmarks
)
else:
warped_image = CPUDeformer.warp_face(
pil_image,
source_landmarks,
target_landmarks
)
output_image = np.array(warped_image).astype(np.float32) / 255.0
output_image = torch.from_numpy(output_image).unsqueeze(0)
return (output_image, processor_settings)
def _convert_to_numpy(self, image):
"""Improved image conversion with channel handling"""
if torch.is_tensor(image):
readme.md
+158
View File
@@ -0,0 +1,158 @@
# Face Processor for ComfyUI
A custom node collection for ComfyUI that provides advanced face detection, alignment, and transformation capabilities using MediaPipe Face Mesh.
## Features
- **Face Detection & Landmark Extraction**: Uses MediaPipe Face Mesh to detect and extract 468 facial landmarks
- **Face Alignment**: Automatic face alignment based on eye positions
- **Face Transformation**:
- Warping between source and target face landmarks
- Scale and translation controls for face shape adjustment
- CPU and CUDA-accelerated processing options
- **Debug Visualization**:
- Visual landmark overlay with customizable parameters
- Support for both detected and target landmark visualization
- Optional landmark labels
## Installation
1. Clone this repository into your ComfyUI's `custom_nodes` directory:
```bash
cd ComfyUI/custom_nodes
git clone [repository-url] face_processor
```
2. Install required dependencies:
```bash
pip install mediapipe opencv-python numpy pandas pillow torch
```
For CUDA acceleration:
1. Install [CUDA Toolkit](https://developer.nvidia.com/cuda-toolkit)
2. Install [CuPy](https://docs.cupy.dev/en/stable/install.html):
```bash
pip install cupy-cuda11x # Replace with your CUDA version
```
## Nodes
### FaceWrapper
Main node for face detection and transformation operations.
#### Inputs:
- `image`: Input image (ComfyUI IMAGE type)
- `mode`: Operating mode
- `Debug`: Visualization of detected landmarks
- `Un-Wrap`: Transform face to normalized position
- `Wrap`: Transform normalized face back to original position
- `device`: Processing device (`CPU` or `CUDA`)
- `show_detection`: Toggle detected landmarks visualization
- `show_target`: Toggle target landmarks visualization
- `landmark_size`: Size of landmark points in visualization
- `show_labels`: Toggle landmark index labels
- `x_scale`: Horizontal scaling factor (0.5 to 1.0)
- `y_transform`: Vertical translation (-0.5 to 0.5)
- `processor_settings`: Optional settings dictionary
#### Outputs:
- `image`: Processed image
- `processor_settings`: Updated settings dictionary
### FaceFitAndRestore
Node for face cropping and restoration operations.
#### Inputs:
- `mode`: Operating mode
- `Fit`: Crop and align face
- `Restore`: Place processed face back in original image
- `image`: Input image
- `padding_percent`: Additional padding around face (0.0 to 1.0)
- `bbox_size`: Output size for cropped face (512, 1024, or 2048)
- `processor_settings`: Required for Restore mode
#### Outputs:
- `image`: Processed image
- `processor_settings`: Updated settings dictionary
- `mask`: Mask indicating face region
## Technical Details
### Core Components
#### Face Detection
- Uses MediaPipe Face Mesh for robust face detection and landmark extraction
- Provides 468 facial landmarks with 3D coordinates
- Supports various input formats (PIL Image, numpy array, torch tensor)
#### Image Processing
- Automatic face rotation based on eye positions
- Aspect ratio-preserving resizing
- Support for square cropping with configurable padding
- Boundary triangulation for complete face warping
#### Face Warping
- Triangle-based warping using predefined mesh topology from MediaPipe Face Mesh
- CPU implementation using pure Python/NumPy
- CUDA-accelerated GPU implementation using CuPy
- Handles both forward and inverse warping
### Performance Considerations
- GPU acceleration requires CUDA toolkit and CuPy
- CPU fallback available for all operations
- Progressive feedback during long operations
- Memory-efficient processing for large images
## Example Usage
Basic face detection and visualization:
```python
face_wrapper = FaceWrapper()
result_image, settings = face_wrapper.detect_face(
image=input_image,
mode="Debug",
device="CPU",
show_detection=True,
show_target=False,
landmark_size=4,
show_labels=True,
x_scale=1.0,
y_transform=0.0
)
```
Face normalization workflow:
1. Detect and normalize face:
```python
# Unwrap face to normalized position
normalized_face, settings = face_wrapper.detect_face(
image=input_image,
mode="Un-Wrap",
device="CUDA",
x_scale=1.0,
y_transform=0.0
)
```
2. Process normalized face with your preferred method
3. Restore face to original position:
```python
# Wrap processed face back
final_image, _ = face_wrapper.detect_face(
image=processed_face,
mode="Wrap",
device="CUDA",
processor_settings=settings
)
```
## License
[Your License Here]
## Acknowledgments
- MediaPipe Face Mesh for facial landmark detection
- ComfyUI project for the node system framework