Atugil-Intelligence/Synas-Detect-1

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

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

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:

@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}}
}