This repository gives a basic layout of the CFSM paper.
The CFSM (Controllable Face Synthesis Model) is a neural network architecture designed for image style transfer. It integrates content encoding, style decoding, and adversarial training through a discriminator to achieve high-quality image generation with diverse styles.
- Image Encoder: Extracts content features from input images.
- Style Decoder: Applies learned styles to the encoded content.
- Discriminator: Evaluates the authenticity of generated images.
- Linear Subspace: Samples styles to guide the generation process.
Make sure you have Python and NumPy installed. You can install NumPy using pip:
pip install numpy
# Usage
##Example Usage
import numpy as np
# Initialize CFSM model
model = CFSM(img_size=112, style_dim=128, n_bases=10)
# Example input image (dummy data)
input_image = np.random.rand(1, 3, 112, 112) # Batch size of 1, RGB image
# Forward pass
output_image, style = model.forward(input_image)
# Compute losses (dummy data for real and fake images)
real_image = np.random.rand(1, 3, 112, 112) # Batch size of 1, RGB image
losses = model.compute_losses(real_image, output_image, style)
# Print losses
print(losses)-img_size (int): Size of the input images (default: 112). -style_dim (int): Dimension of the style representation (default: 128). -n_bases (int): Number of bases for the linear subspace (default: 10). -l_a, u_a (float): Lower and upper bounds for the style norms (default: 0, 6). -l_m, u_m (float): Lower and upper margins for the identity loss (default: 0.05, 0.65).
- Purpose: Encodes input images to extract content features.
- Purpose: Decodes the content features and applies styles to generate output images.
- Purpose: Discriminates between real and generated images to train the generator.
- Purpose: Samples styles from a learned linear subspace to guide the style generation.
- Purpose: A basic fully connected network used in the style decoder.