tgallup/city-gan-2

Progressive GAN Image Generator for Cities

This model was trained on a custom city image dataset using a Progressive GAN architecture with techniques to prevent mode collapse.

Training Information

• Epochs: 2400/10000
• Current Date: 2025-05-10 22:59:12
• Current Resolution: 1x1
• Alpha: 1.0
• Latent Dimension: 100
• Batch Size: 64

Latest Training Metrics

• Generator Loss: 0.7387560606002808
• Discriminator Loss: 0.6835118532180786
• Discriminator Accuracy: 0.5859375

Sample Images

Training Techniques

• Progressive Growing from 8x8 to 64x64
• Minibatch Discrimination to prevent mode collapse
• Instance Noise for training stability

Usage

import torch
from torch import nn

# Define generator architecture
class Generator(nn.Module):
    # ... (architecture code)

# Load the model
generator = Generator(latent_dim=100, channels=3)
generator.load_state_dict(torch.load('model/generator_epoch_2400.pt'))
generator.eval()

# Generate images
z = torch.randn(1, 100)
with torch.no_grad():
    fake_image = generator(z)
    # Convert to range [0, 1] for display
    fake_image = (fake_image + 1) / 2.0

Training Progress

Training is currently 24.0% complete. Current resolution phase: 1 of 4