Mitchins/image-medium-classifier-efficientnetv2-s-v1

Anime/Real/Rendered Image Classifier (TF-EfficientNetV2-S)

Higher-capacity classifier with improved generalization for anime, photo, and 3D detection.

Model Details

• Architecture: TF-EfficientNetV2-S (timm)
• Input Size: 224×224 RGB
• Classes: anime, real, rendered
• Parameters: 21.5M (4× larger than B0)
• Validation Accuracy: 97.55% (+0.11% vs B0)
• Training Speed: ~3 min/epoch (GPU)
• Inference Speed: ~60ms per image (RTX 3060)

Performance

Class	Precision	Recall	F1-Score
anime	1.00	0.97	0.98
real	0.98	0.99	0.98
rendered	0.93	0.90	0.91
macro avg	0.97	0.95	0.96

Comparison to EfficientNet-B0

Metric	B0	V2-S	Winner
Final Accuracy	97.44%	97.55%	V2-S +0.11%
Best Accuracy	97.99%	97.99%	Tied
Params	5.3M	21.5M	B0 (lighter)
Speed	1 min/epoch	3 min/epoch	B0 (faster)
Convergence	Epoch 4	Epoch 13	B0 (faster)

Verdict: V2-S learns training data better with marginally improved generalization. Use B0 for speed, V2-S for accuracy.

Usage

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

# Load model
model = timm.create_model('tf_efficientnetv2_s', 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]),
])

image = Image.open('image.jpg').convert('RGB')
x = transform(image).unsqueeze(0)

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

labels = ['anime', 'real', 'rendered']
print(f"{labels[pred_class]}: {probs[0, pred_class]:.2%}")

Dataset

• Real: 5,000 COCO 2017 validation images
• Anime: 2,357 curated animation frames
• Rendered: 1,610 AAA games + 61 Pixar stills
• Total: 8,967 images (8,070 train / 897 perceptually-hashed val)

Training Details

• Augmentation: None (raw resize to 224×224)
• Optimizer: AdamW (lr=0.001)
• Loss: CrossEntropyLoss with class weighting
• Epochs: 20
• Batch Size: 40 (GPU memory constrained)
• Hardware: NVIDIA RTX 3060 (12GB)

Known Behavior

• Better Anime Detection: Perfect precision (1.00) but 97% recall
• Stronger Real Recognition: 99% recall on real images
• Rendered Uncertainty: 90% recall suggests photorealistic games still challenging
• Slower Inference: ~3× slower than B0 due to model size

Recommendations

• Production: Ensemble both models for maximum confidence
• Real-time: Use B0 for speed-critical applications
• Accuracy-critical: Use V2-S as primary model
• Confidence Thresholding: Only trust predictions >80% confidence

License

OpenRAIL - Free for research and educational purposes