ntkuhn/anime-score-model

Anime Face Score Matching Model

A score-based generative model trained to generate 64x64 anime-style faces using Denoising Score Matching.

Model Details

• Model type: NCSN / Score Matching
• Training data: Anime faces dataset
• Image size: 64x64 RGB
• Sigma: 0.15
• Architecture: Improved U-Net

Usage

import torch
# Load your model class definitions here

# Load the model
checkpoint = torch.load("model.pt")
model = ImprovedScoreNetwork(in_channels=3)
model.load_state_dict(checkpoint['ema_model_state_dict'])  # Use EMA weights
model.eval()

# Generate samples using Langevin Dynamics
# (See sample_langevin function in training code)