model_card.md

---
license: openrail
language: en
library_name: timm
tags:
  - image-classification
  - anime
  - real
  - rendered
  - 3d-graphics
datasets:
  - coco
  - custom-anime
  - steam-screenshots
---

# EfficientNet-B0 - Anime/Real/Rendered Classifier

Fast, lightweight image classifier distinguishing photographs from anime and 3D rendered images.

## Model Summary

- **Model Name:** efficientnet_b0
- **Framework:** PyTorch + TIMM
- **Input:** 224×224 RGB images
- **Output:** 3 classes (anime, real, rendered)
- **Parameters:** 5.3M
- **Size:** 16.2 MB

## Intended Use

Classify images into three categories:
- **anime**: Drawn 2D or cel-shaded animation
- **real**: Photographs and real-world footage
- **rendered**: 3D graphics (games, CGI, Pixar, etc.)

## Performance

**Validation Accuracy:** 97.44%

| Class | Precision | Recall | F1-Score | Support |
|-------|-----------|--------|----------|---------|
| anime | 0.98 | 0.99 | 0.99 | 236 |
| real | 0.98 | 0.98 | 0.98 | 500 |
| rendered | 0.96 | 0.93 | 0.94 | 161 |
| **weighted avg** | **0.97** | **0.97** | **0.97** | **897** |

## Training Data

- **Real images:** 5,000 COCO 2017 validation set images
- **Anime images:** 2,357 curated animation frames and key scenes
- **Rendered images:** 1,549 AAA game screenshots (Metacritic ≥75) + 61 Pixar movie stills
- **Total:** 8,967 images, 8,070 training, 897 validation (perceptually-hashed for diversity)

## Training Details

- **Framework:** PyTorch
- **Augmentation:** Resize only (224×224)
- **Loss Function:** CrossEntropyLoss with inverse frequency class weights
- **Optimizer:** AdamW (lr=0.001)
- **Batch Size:** 80
- **Epochs:** 20
- **Hardware:** NVIDIA RTX 3060 (12GB VRAM)
- **Training Time:** ~20 minutes

## Limitations

1. Photorealistic video games sometimes classified as real (90% recall on rendered class)
2. Cel-shaded games may score as anime rather than rendered
3. Artistic 3D renders (Pixar, high-quality CGI) show mixed confidence
4. Performance degrades on images <224×224

## Recommendations

- Use confidence threshold of ≥80% for reliable predictions
- For critical applications, ensemble with tf_efficientnetv2_s
- Check confusion patterns in own use cases
- Manually review edge cases (game screenshots, stylized renders)

## How to Use

```python
from PIL import Image
import torch
from torchvision import transforms
import timm
from safetensors.torch import load_file

# Load
model = timm.create_model('efficientnet_b0', num_classes=3, pretrained=False)
state_dict = load_file('model.safetensors')
model.load_state_dict(state_dict)
model.eval()

# Prepare image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
img = Image.open('image.jpg').convert('RGB')
x = transform(img).unsqueeze(0)

# Infer
with torch.no_grad():
    logits = model(x)
    probs = torch.softmax(logits, dim=1)
    pred = probs.argmax().item()

labels = ['anime', 'real', 'rendered']
print(f"{labels[pred]}: {probs[0, pred]:.1%}")
```

## Benchmarks

**Inference Speed (RTX 3060)**
- Single image: ~20ms
- Batch of 32: ~150ms

**Accuracy Comparison**
| Model | Accuracy | Speed | Params |
|-------|----------|-------|--------|
| EfficientNet-B0 | 97.44% | Fast | 5.3M |
| TF-EfficientNetV2-S | 97.55% | Moderate | 21.5M |

## Ethical Considerations

This model classifies images by visual style/source. Potential misuse:
- Detecting deepfakes/AI-generated content (not designed for this)
- Filtering user-generated content (may have cultural bias)
- Surveillance or profiling

**Recommendations:**
- Use with human review for content moderation
- Test on your target domain before deployment
- Don't rely solely on automatic classification for safety-critical decisions
- Consider cultural representation in anime/rendered content

## Contact

For questions or issues: [GitHub repo]

## License

OpenRAIL (Open Responsible AI License) - free for research and commercial use with proper attribution