README.md

---
license: mit
base_model: openai/clip-vit-large-patch14
tags:
  - ai-detection
  - deepfake-detection
  - vision
  - image-classification
  - clip
  - synthetic-image-detection
pipeline_tag: image-classification
library_name: transformers
metrics:
  - accuracy
model-index:
  - name: Synas Detect 1
    results:
      - task:
          type: image-classification
          name: Synthetic vs Real Image Classification
        dataset:
          type: ComplexDataLab/OpenFake
          name: OpenFake Core Dataset
        metrics:
          - type: accuracy
            value: 95.0
---

# Synas Detect 1

**Synas Detect 1** is a high-accuracy vision classifier (~304 million parameters) developed by **Atugil Intelligence**. Fine-tuned on modern synthetic media, it specializes in determining whether an image is an authentic photograph or AI-generated by models such as **Midjourney**, **Stable Diffusion**, **DALL·E**, **Imagen**, and similar generative systems.

---

# Model Overview

| Attribute | Detail |
| :--- | :--- |
| **Developed by** | Atugil Intelligence |
| **Model Name** | Synas Detect 1 |
| **Model Type** | Vision Transformer / Image Classifier |
| **Base Architecture** | `openai/clip-vit-large-patch14` |
| **Parameter Count** | ~304 Million |
| **License** | MIT |
| **Primary Task** | Binary Image Classification (`0`: AI Generated / Fake, `1`: Real Photograph) |

---

# Intended Use

## Primary Use Cases

- **Media Authenticity Verification** – Detect synthetic imagery for journalism, publishing, and digital archives.
- **Content Moderation** – Flag AI-generated uploads within trust and safety pipelines.
- **Forensic Visual Analysis** – Assist researchers and investigators by identifying subtle generation artefacts.

## Limitations

While Synas Detect 1 performs strongly across a wide variety of datasets, it has several limitations:

- Extremely compressed JPEG images or aggressive social media recompression may reduce confidence.
- Heavy cropping, resizing, or post-processing can remove useful forensic signals.
- The model is not explicitly trained against adversarial perturbations or deliberate evasion attacks.
- Classification confidence should not be treated as absolute proof of authenticity.

---

# Training Setup

| Setting | Value |
| :--- | :--- |
| **Dataset** | `ComplexDataLab/OpenFake` (Core Partition) |
| **Hardware** | 2× NVIDIA T4 GPUs |
| **Precision** | Mixed Precision (FP16) |
| **Framework** | PyTorch + Transformers |
| **Parallelism** | DataParallel |
| **Effective Batch Size** | 64 |
| **Learning Rate** | 2e-5 |
| **Optimizer** | AdamW + Linear Learning Rate Decay |

The model was fine-tuned on high-quality synthetic and real image pairs while preserving the strong visual representations learned by CLIP.

---

# Performance

| Metric | Score |
| :--- | :--- |
| **In-Distribution Accuracy** | ~96% |
| **Out-of-Sample Accuracy** | ~92–95% |

Performance varies depending on image quality, compression level, and the generation model used.

---

# Usage

## Transformers Pipeline

```python
from transformers import pipeline
from PIL import Image

classifier = pipeline(
    "image-classification",
    model="Atugil-Intelligence/synas-detect-1"
)

image = Image.open("path_to_image.jpg").convert("RGB")

predictions = classifier(image)

print("Synas Detect 1 Predictions")
print("-" * 40)

for prediction in predictions:
    print(
        f"{prediction['label']} "
        f"({prediction['score'] * 100:.2f}%)"
    )
```

---

## Native PyTorch Implementation

```python
import torch
from PIL import Image
from transformers import (
    AutoImageProcessor,
    AutoModelForImageClassification,
)

MODEL_ID = "Atugil-Intelligence/synas-detect-1"

device = torch.device(
    "cuda" if torch.cuda.is_available() else "cpu"
)

processor = AutoImageProcessor.from_pretrained(MODEL_ID)
model = AutoModelForImageClassification.from_pretrained(
    MODEL_ID
).to(device)

model.eval()

image = Image.open("path_to_image.jpg").convert("RGB")

inputs = processor(
    image,
    return_tensors="pt"
).to(device)

with torch.no_grad():
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)
    prediction = probabilities.argmax(dim=-1).item()
    confidence = probabilities[0][prediction].item()

labels = {
    0: "AI Generated (Fake)",
    1: "Real Photograph",
}

print(f"Prediction : {labels[prediction]}")
print(f"Confidence : {confidence * 100:.2f}%")
```

---

# License

Synas Detect 1 is released under the **MIT License**.

## Attribution

This model is fine-tuned from OpenAI's **CLIP ViT-Large Patch-14** vision backbone.

**Copyright**

- Copyright © 2021 OpenAI (Base Model)
- Copyright © 2026 Atugil Intelligence (Fine-Tuned Weights)

---

# Citation

If you use **Synas Detect 1** in research, benchmarks, or production systems, please cite:

```bibtex
@misc{synas_detect_1_2026,
  author       = {Atugil Intelligence},
  title        = {Synas Detect 1: High-Accuracy Synthetic Image Classifier},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Atugil-Intelligence/synas-detect-1}}
}
```