🎨 AnimeFace-DDPM
A Denoising Diffusion Probabilistic Model (DDPM) implemented from scratch in PyTorch for unconditional anime face generation.
The model was trained on the Anime Face Dataset in a Kaggle Notebook using an NVIDIA Tesla T4 GPU. During inference, Exponential Moving Average (EMA) shadow weights together with DDIM sampling are used to generate higher-quality images with significantly fewer sampling steps.
Model Details
| Property | Value |
|---|---|
| Model | DDPM |
| Framework | PyTorch |
| Architecture | Custom U-Net |
| Parameters | 7.52 Million |
| Dataset | Anime Face Dataset |
| Image Resolution | 64X64 |
| Diffusion Timesteps | 1000 |
| Sampling | DDIM |
| EMA | Yes |
| Optimizer | AdamW |
Available Checkpoints
This repository contains two checkpoints.
ddpm.pth
The standard model weights obtained directly after training.
ema.pth
The Exponential Moving Average (EMA) shadow weights.
These weights are recommended for inference because they generally produce sharper and more stable image generations.
Training
The model was trained entirely from scratch using PyTorch.
Training Configuration
- Dataset: Anime Face Dataset
- Training Platform: Kaggle
- GPU: NVIDIA Tesla T4
- Optimizer: AdamW
- Diffusion Timesteps: 1000
- EMA enabled during training
- DDIM used during inference
Intended Use
This model is intended for
- Learning diffusion models
- Educational purposes
- Research
- Anime image generation
- Experimenting with DDPMs
Limitations
- Generates only anime-style faces.
- Performance is limited to the distribution of the training dataset.
- This is an unconditional diffusion model and cannot generate images from text prompts.
Citation
If you use this model in your work, please cite the original diffusion papers.
@article{ho2020ddpm,
title={Denoising Diffusion Probabilistic Models},
author={Jonathan Ho and others},
year={2020}
}
@article{song2020ddim,
title={Denoising Diffusion Implicit Models},
author={Jiaming Song and others},
year={2020}
}
Acknowledgements
- PyTorch
- Hugging Face
- Kaggle
- DDPM (Ho et al., 2020)
- DDIM (Song et al., 2020)
License
This project is released under the MIT License.