Nininkkka/DCGAN-winter-landscapes-64x64

DCGAN Winter Landscape Generator (64x64)

A DCGAN trained from scratch on a custom 64x64 winter landscape dataset. The model generates unique scenery including daytime landscapes, dusk, and night scenes with realistic sky gradients.

The training dataset is owned by the author and is licensed under CC-BY-NC-4.0.

Model Architecture

• —Generator: ConvTranspose2d, BatchNorm, ReLU, ending with Tanh.
• —Discriminator: LeakyReLU, BatchNorm, Sigmoid output.
• —Latent Dimension: 100
• —Resolution: 64x64

Training & Evaluation

Trained for 100 epochs using PyTorch:

• —Optimizer: Adam (beta1 = 0.5)
• —Batch Size: 128
• —Learning Rate: Dynamic scaling (lr = lr * (batch_size / 32))
• —Label Smoothing: 0.9 for real labels.

Final Losses (Epoch 100): | Metric | Score | | :--- | :--- | | Discriminator Loss | 0.2978 | | Generator Loss | 2.9600 |

Interactive Demo

Try it in your browser: Gradio Space [image]