README.md

---
license: mit
tags:
- pytorch
- diffusers
- unconditional-image-generation
- diffusion-models
- anime
- anime-faces
- ddpm
---

# Anime Face Diffusion Model 🎨

A fine-tuned diffusion model for generating high-quality anime faces using DDPM. This model is based on Google's pre-trained `ddpm-celebahq-256` model and fine-tuned on 7,000+ anime face images.

## Model Details

- **Model Type**: Denoising Diffusion Probabilistic Model (DDPM)
- **Base Model**: [google/ddpm-celebahq-256](https://huggingface.co/google/ddpm-celebahq-256)
- **Task**: Unconditional Image Generation (256×256 anime faces)
- **Training Data**: 7,000+ high-quality anime face images
- **Framework**: 🧨 Diffusers
- **License**: MIT

## Training Parameters

- **Learning Rate**: 2e-5
- **Epochs**: 15
- **Batch Size**: 4
- **Gradient Accumulation Steps**: 2
- **Training Steps**: ~26,250 (1750 steps/epoch × 15 epochs)
- **Optimizer**: AdamW
- **Loss**: MSE (Mean Squared Error)

## Usage

### Basic Usage

```python
from diffusers import DDPMPipeline
import torch

# Load the model
pipeline = DDPMPipeline.from_pretrained("abcd2019/Anime-face-generation")
device = "cuda" if torch.cuda.is_available() else "cpu"
pipeline = pipeline.to(device)

# Generate a single image
image = pipeline(num_inference_steps=100).images[0]
image.save("anime_face.png")
```

### Generate Multiple Images

```python
from diffusers import DDPMPipeline

pipeline = DDPMPipeline.from_pretrained("abcd2019/Anime-face-generation")
pipeline = pipeline.to("cuda")

# Generate 5 anime faces
images = pipeline(batch_size=5, num_inference_steps=100).images

for i, image in enumerate(images):
    image.save(f"anime_face_{i}.png")
```

### Adjust Inference Steps for Quality vs Speed

```python
# Fast generation (fewer steps, less quality)
fast_image = pipeline(num_inference_steps=50).images[0]

# High quality (more steps, slower)
quality_image = pipeline(num_inference_steps=150).images[0]

# Recommended: 100 steps for good balance
balanced_image = pipeline(num_inference_steps=100).images[0]
```

### Use Different Scheduler

```python
from diffusers import DDPMPipeline, DDIMScheduler

pipeline = DDPMPipeline.from_pretrained("abcd2019/Anime-face-generation")

# Switch to DDIM for faster sampling
scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
scheduler.set_timesteps(num_inference_steps=50)
pipeline.scheduler = scheduler

fast_image = pipeline().images[0]  # Generates in ~50 steps instead of 1000
```

## Model Performance

- **Training Loss**: ~0.0077 (final epoch)
- **Image Resolution**: 256×256 pixels
- **Inference Speed**: ~30-60 seconds per image (depending on steps)
- **Recommended Inference Steps**: 100 (for best quality)
- **Generated Face Styles**: Wide diversity of anime faces with various:
  - Hair colors and styles
  - Eye colors and expressions
  - Face shapes and features
  - Skin tones

## Limitations & Bias

- **Resolution**: Limited to 256×256 pixels (inherent to model architecture)
- **Style**: Specifically trained on anime faces, may not generate realistic/photorealistic faces
- **Diversity**: Generated faces are limited to patterns in training data
- **Quality Variation**: Face shape clarity depends on inference steps (higher = better)

## Training Details

### Data Preparation
- **Dataset**: Anime Face Dataset (Kaggle)
- **Total Images**: 7,000
- **Selection Method**: Top quality images by file size
- **Preprocessing**: 
  - Resized to 256×256
  - Random horizontal flip (50% probability)
  - Normalized to [-1, 1]

### Fine-tuning Approach
- Started from pre-trained `ddpm-celebahq-256`
- Fine-tuned with low learning rate to preserve general face generation knowledge
- Adapted to anime-specific features (large eyes, stylized features, etc.)

### Training Dynamics
- **Epoch 0-3**: Model adapts from photorealistic to anime style
- **Epoch 4-8**: Loss continues to decrease, anime features solidify
- **Epoch 9+**: Marginal improvements, risk of overfitting

## Ethical Considerations

This model generates synthetic anime faces and should not be used to:
- Create misleading/deceptive content
- Generate non-consensual images of real people
- Violate any local laws or regulations

## Recommended Citation

If you use this model in your research or project, please credit:
- The original DDPM paper
- Google's pre-trained `ddpm-celebahq-256` model
- This fine-tuned adaptation

## Future Improvements

Potential enhancements for future versions:
- Higher resolution (512×512 or more)
- Conditional generation (text-to-image for anime faces)
- Better diversity through larger training datasets
- Improved training with advanced schedulers or techniques

## Resources

- 📚 [Diffusion Models Class](https://github.com/huggingface/diffusion-models-class)
- 📖 [Diffusers Documentation](https://huggingface.co/docs/diffusers)
- 📄 [DDPM Paper](https://arxiv.org/abs/2006.11239)
- 🤗 [Hugging Face Hub](https://huggingface.co)

---

**Created**: 2025-12-28

**Model Card Contact**: [Your Name/Username]