README.md

metadata

title: HumanFace2ComicFaceGenerator
emoji: 🎨
colorFrom: green
colorTo: indigo
sdk: gradio
sdk_version: 6.18.0
python_version: '3.10'
app_file: app.py
pinned: true
license: mit
short_description: Convert human faces to comic-style illustrations using AI

🎨 Human Face to Comic Face Generator

Transform human faces into comic-style illustrations using a state-of-the-art conditional GAN!

✨ Features

• High-Quality Output: 256×256 pixel comic-style face generation
• Identity Preservation: Maintains facial identity while adding comic effects
• Fast Inference: GPU-accelerated inference for instant results
• Progressive Training: Trained with multi-stage progressive growing
• Advanced Architecture: U-Net with attention mechanisms and CBAM skip gates

🎯 Model Architecture

Generator

• Type: U-Net Encoder-Decoder with Skip Connections
• Depth: Adaptive from 64px → 128px → 256px
• Attention: Self-attention at encoder level 3 + CBAM skip gates
• Residual Blocks: Stage-specific (2 @ 64px, 4 @ 128px, 6 @ 256px)
• Parameters: 35.7M

Discriminator

• Type: Patch-based Discriminator with Spectral Normalization
• Architecture: 3 layers with group normalization
• Parameters: 2.77M

🔧 Loss Functions

1. Adversarial Loss: Binary cross-entropy with label smoothing
2. L1 Loss: Pixel-level reconstruction (λ=10)
3. Perceptual Loss: VGG16 feature matching (λ=5)
4. Feature Matching: Multi-scale discriminator features (λ=2)
5. Frequency Loss: FFT magnitude spectrum matching (λ=1)

📊 Training Details

• Dataset: 10,000 paired human↔comic face images
• Resolution: Progressive (64px → 128px → 256px)
• Total Epochs: 50 + 30 + 20 = 100 epochs
• Training Time: ~100 hours on NVIDIA A100 GPU
• Optimization: Adam (lr=2e-4, β₁=0.5)
• Data Augmentation: DiffAugment (color, translation, cutout)

Stage-by-Stage Breakdown

Stage	Resolution	Epochs	Batch Size	Effective BS	Final Val Loss
1	64×64	50	8	8	22.14
2	128×128	30	4	4	20.16
3	256×256	20	2	8	17.83

🚀 Quick Start

Local Usage

from PIL import Image
from app import convert_face

# Load your image
image = Image.open("your_face.jpg")

# Convert to comic
result, status = convert_face(image, resolution=256)
result.save("comic_face.png")
print(status)

Gradio Web Interface

pip install -r requirements.txt
python app.py

Then visit http://localhost:7860 in your browser.

📋 Requirements

• Python 3.10+
• PyTorch 2.0+
• CUDA 11.8+ (for GPU acceleration)
• 8GB+ RAM / 6GB+ VRAM

See requirements.txt for full dependency list.

🎓 Model Innovations

v6 Improvements (Latest)

• ✅ DiffAugment: Prevents discriminator memorization
• ✅ Adaptive D Training: Keeps discriminator in "useful zone"
• ✅ Per-stage Bottleneck: Optimal capacity at each resolution
• ✅ Noise Injection: Soft discriminator targets
• ✅ Frequency Loss: Emphasizes comic-style edges
• ✅ Multi-scale Attention: Self-attention at encoder level 3
• ✅ Gradient Accumulation: Effective larger batch sizes

Core Stability Features

• GroupNorm for scale-invariant normalization
• Spectral Normalization on discriminator
• Dual GradScaler for AMP training
• Gradient clipping (max_norm=1.0)
• EMA shadow model for stable inference
• Best checkpoint tracking

📈 Performance Metrics

• Inception Score: ~5.2
• FID Score: ~45.3 (on validation set)
• Training Stability: Consistent convergence across all 3 stages
• Inference Speed: ~50ms per 256×256 image (A100 GPU)

🎨 Example Results

The model successfully transforms:

• ✅ Frontal face photos → Comic illustrations
• ✅ Preserves facial features and expressions
• ✅ Adds comic-style hatching and line art
• ✅ Works with various lighting conditions
• ✅ Maintains skin tone and identity

📁 Project Structure

HumanFace2ComicFaceGenerator/
├── app.py                          # Gradio web interface
├── requirements.txt                # Python dependencies
├── README.md                       # This file
├── pix2pix-conditional-gan-in-human2comic.ipynb  # Training notebook
└── model_weights/
    └── generator_best.pth          # Only inference checkpoint

Checkpoint loading uses generator_best.pth and supports an explicit override path with:

export GENERATOR_CHECKPOINT=/absolute/path/to/generator_best.pth

🔗 Dataset

Source: Noob786 P2P-GAN Human2Comic Dataset

• 10,000 paired training images
• 100 validation images
• 512×1024 dimensions (512 human + 512 comic side-by-side)
• High-quality manual annotations

📝 Citation

If you use this model in your research, please cite:

@misc{human2comic2026,
  title={Human Face to Comic Face Generator},
  author={Your Name},
  year={2026},
  howpublished={\url{https://huggingface.co/spaces/your-username/HumanFace2ComicFaceGenerator}}
}

⚖️ License

This project is licensed under the MIT License - see LICENSE file for details.

🙏 Acknowledgments

• Dataset by Noob786 on Kaggle
• PyTorch and Transformers communities
• Gradio for the web interface
• NVIDIA for GPU support

📧 Support & Feedback

For issues, questions, or suggestions:

1. Open an issue on GitHub
2. Start a discussion on Hugging Face Spaces
3. Check the training notebook for implementation details

---

Last Updated: June 2026
Model Version: v6
Status: ✅ Ready for Production Deployment