beatrizfarias/pokemon-gan

Unconditional DCGAN — Pokemon Image Generator

Model Description

A DCGAN-style unconditional Generative Adversarial Network trained to generate Pokemon-style images from random noise.

The Generator uses transposed convolutions to progressively upsample a latent vector into a 64×64 RGB image, while the Discriminator uses strided convolutions to downsample and classify images as real or fake.

Architecture

Generator

• Input: noise vector z (latent dim = 100)
• Linear projection: 100 → 512×4×4, reshaped to (512, 4, 4)
• Transposed conv layers: 4×4 → 8×8 → 16×16 → 32×32 → 64×64
• BatchNorm + ReLU activations, Tanh output
• Output: (3, 64, 64) RGB image

Discriminator

• Input: (3, 64, 64) RGB image
• Strided conv layers: 64×64 → 32×32 → 16×16 → 8×8 → 4×4
• BatchNorm + LeakyReLU activations, Sigmoid output

Training

• Dataset: huggan/pokemon (7,357 images)
• Epochs: 200
• Batch size: 64
• Optimizer: Adam — Generator lr = 0.0002, Discriminator lr = 0.0001 (β = 0.5, 0.999)
• Loss: Binary Cross-Entropy with label smoothing (real targets = 0.9)
• Generator steps per batch: 2 (to balance training against the Discriminator)
• Augmentation: random crop, horizontal flip, color jitter
• Hardware: Apple MPS (Metal Performance Shaders)

Usage

from huggingface_hub import hf_hub_download
import torch
from pokemon_gan_model import Generator

model = Generator(latent_dim=100)
weights_path = hf_hub_download(repo_id="beatrizfarias/pokemon-gan", filename="pokemon_gan_generator.pth")
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

z = torch.randn(1, 100)
with torch.no_grad():
    img = model(z)  # shape: (1, 3, 64, 64)

Results

After 200 epochs, the generator produces images with clear creature-like silhouettes and Pokemon-style color palettes on white backgrounds.