AI Image Detectors (Ensemble Weights)

This repository contains trained PyTorch model weights used for detecting AI-generated images vs real photographs.

The models are designed to be used together as an ensemble and are consumed by the Hugging Face Space:

👉 https://huggingface.co/spaces/divine2k/ai_image-classification

Models Included

Model	Architecture	File
Model 1	ResNet50	`resnet50_ai_real_final.pth`
Model 2	EfficientNet-B0	`efficientNet_BO_Final.pth`
Model 3	ConvNeXt Tiny	`convNext_final.pth`

All models are binary classifiers:

Output < 0.45 → AI-generated
Output 0.45–0.60 → Uncertain
Output > 0.60 → Real

Final prediction is obtained via majority voting.

Training Overview

Task: AI-generated image detection
Classes: AI, Real
Training data size: ~30,000 high-resolution images
Generators included:
- Stable Diffusion (1.5, SDXL)
- Midjourney
- DALL·E
- Grok (limited)
Fine-tuning strategy:
1. Train classifier head
2. Unfreeze final backbone layers

Intended Use

✅ Research
✅ Demonstrations
✅ Academic projects
✅ Detection benchmarking

❌ Not intended for forensic or legal decisions
❌ Not guaranteed to generalize to unseen generators or heavy post-processing

How to Use (PyTorch)

import torch
import torch.nn as nn
from torchvision.models import resnet50
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="divine2k/ai-image-detectors",
    filename="resnet50_ai_real_final.pth"
)

model = resnet50(weights=None)
model.fc = nn.Linear(model.fc.in_features, 1)
model.load_state_dict(torch.load(path, map_location="cpu"))
model.eval()

⚠️ Important: These files contain state_dict only. You must recreate the architecture before loading.

Limitations

Performance drops on:
- Heavily compressed images
- Strong post-processing (filters, resizing, screenshots)
- New generators unseen during training
Thresholds may require calibration per dataset

Ethical Considerations

This model is provided to support:

Transparency in synthetic media
Research into AI image detection

It should not be used to:

Make accusations
Enforce moderation without human review
Claim absolute authenticity

License

MIT License You are free to use, modify, and redistribute with attribution.

Citation

If you use this work in research or demos, please reference:

divine2k/ai-image-detectors

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

divine2k
/

ai-image-detectors