gabehubner/vae-256px-32z

UNet-Style VAE for 256x256 Image Reconstruction

This model is a UNet-style Variational Autoencoder (VAE) trained on the CC3M dataset for high-quality image reconstruction and generation. It integrates adversarial, perceptual, and identity-preserving loss terms to improve semantic and visual fidelity.

Architecture

• Encoder/Decoder: Multi-scale UNet architecture
• Latent Space: 32-channel latent bottleneck with reparameterization (mu, logvar)
• Losses:
  • L1 reconstruction loss
  • KL divergence with annealing
  • LPIPS perceptual loss (VGG backbone)
  • Identity loss via MoCo-v2 embeddings
  • Adversarial loss via Patch Discriminator w/ Spectral Norm

$${\mathcal{L}}_{total}={\mathcal{L}}_{recon}+{\mathcal{L}}_{PIPS}+0.5\ast {\mathcal{L}}_{GAN}+0.1\ast {\mathcal{L}}_{ID}+{10}^{-6}\ast {\mathcal{L}}_{KL}\backslash mathcal\{L\}\_\{total\}\; =\; \backslash mathcal\{L\}\_\{recon\}\; +\; \backslash mathcal\{L\}\_\{PIPS\}\; +\; 0.5\; *\; \backslash mathcal\{L\}\_\{GAN\}\; +\; 0.1\; *\backslash mathcal\{L\}\_\{ID\}\; +\; 10^\{-6\}\; *\backslash mathcal\{L\}\_\{KL\}$$

Reconstructions

Input	Output

Training Config

Hyperparameter	Value
Dataset	CC3M (850k images)
Image Resolution	256 x 256
Batch Size	16
Optimizer	AdamW
Learning Rate	5e-5
Precision	bf16 (mixed precision)
Total Steps	210,000
GAN Start Step	50,000
KL Annealing	Yes (10% of training)
Augmentations	Crop, flip, jitter, blur, rotation

Trained using a cosine learning rate schedule with gradient clipping and automatic mixed precision (torch.cuda.amp)

Note: Model class will be available for usage once original repository becomes open-source (W.I.P)