pokemon-gan

Pokémon GAN (course project)

Unconditional generative models for small 64×64 RGB Pokémon-style sprites. This repository hosts generator weights from a concrete training run.

Files in this model repo

Trained weights (FastGANLiteGenerator)

The primary artifact is a FastGANLiteGenerator trained with python -m pokemon_gan.train (FastGAN-lite generator + spectrally-normalized DCGAN-style discriminator, hinge loss by default). Checkpoint step 35000 (end of epoch 648).

Input noise: (B, nz, 1, 1) with nz=100 (default training config).

Output: (B, 3, 64, 64) in [-1, 1] (tanh).

Load the generator (example)

import torch
from huggingface_hub import hf_hub_download

# From a checkout of the week-06 lab repo (includes pokemon_gan.fastgan).
import sys
sys.path.insert(0, "/path/to/week-06")
from pokemon_gan.fastgan import FastGANLiteGenerator

repo_id = "<YOUR_REPO_ID>"  # e.g. username/pokemon-gan
weights = hf_hub_download(repo_id, "generator_step_00035000.pth")

G = FastGANLiteGenerator(nc=3, nz=100, ngf=64, image_size=64)
G.load_state_dict(torch.load(weights, map_location="cpu", weights_only=True))
G.eval()
with torch.no_grad():
    z = torch.randn(4, 100, 1, 1)
    fake = G(z)

Training context (defaults)

Typical TrainConfig defaults from pokemon_gan/train.py: batch size 128, lr_g=1e-4, lr_d=4e-4, Adam (beta1=0, beta2=0.999), hinge adversarial loss, image size 64.

Limitations

• Small dataset setting; samples are stylistic and may show mode collapse or artifacts.
• No warranty of biological or franchise accuracy for generated sprites.

References

• Han Zhang et al., Self-Attention Generative Adversarial Networks, ICML 2019.
• FastGAN: Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis, ICLR 2021 (this repo uses a lite DCGAN+SE variant in fastgan.py, not a full FastGAN reproduction).