- U-Net Discriminator: Skip connections enable multi-scale feedback for fine facial details
- L1 + Perceptual Loss: Combines edge preservation (L1) and semantic coherence (VGG19)
- 25-Epoch Training: Achieves stable results on CelebA (64×64)
- VAE Encoder: Ensures that the image is not derived from a purely random latent dimension but instead encapsulates probabilistic features that help maintain consistency with the training image dataset.
- Decoder: Reconstructs images from latent vectors through upsampling layers and tanh activation, trained with combined L1 and perceptual losses to preserve sharp features and semantic consistency.
- U-Net Discriminator: Integrates both a global score for distinguishing between real and fake samples, and a PatchGAN output at the final layer with sigmoid activation which help capture both local and global features effectively.
- Balanced Losses: Ensures high-level semantic consistency
- Adversarial loss (GAN)
- L1 reconstruction (λ=100)
- Perceptual loss (λ=0.3)
-
Original: These serve as reference for evaluating the model’s ability to preserve identity, texture, and structure.
-
Reconstructed: Displays reconstructions generated by encoding and decoding the test images using the trained VAE-GAN.
-
Generated Samples: Displays a grid of images generated by sampling random noise vectors from a standard Gaussian distribution & decodes them via generator
