boluobobo/ItsNotAI-ai-detector-v1

ItsNotAI v1 - Multiclass AI Image Detector

Detect AI-generated images | Identify the AI generator | Verify human-made artwork

A Vision Transformer model that detects AI-generated images and identifies the specific AI generator used (33 classes).

Website: https://itsnotai.org

---

Newer Version Available

Version	Features	Link
v2 (Latest)	Dual-head, FLUX detection, improved Midjourney	ItsNotAI-ai-detector-v2
v1 (This)	Single-head, 33-class classification	Current page

Recommendation: Use v2 for better binary (Real/AI) classification. Use v1 if you only need source identification without the binary head.

---

Note: This is one of the models used by ItsNotAI. For official verification at itsnotai.org, we use an ensemble of multiple models combined with human expert review to ensure maximum accuracy.

---

About ItsNotAI

Most AI detectors focus on catching AI usage. ItsNotAI takes the opposite approach: helping artists prove their work is human-made.

Key Features

• Verifiable Label: Beyond just a percentage score, we provide artists with a verifiable "Not AI" label that can be embedded in their work.

• Industry-Focused: We specialize in digital painting, manga illustration, and texture design, developed in deep collaboration with 100+ professional artists.

• Artist-First: Our industry endorsements and artist partnerships create a trust network that goes beyond pure technical metrics.

Use Cases

• Artists & Creators: Prove your artwork is human-made, protect your reputation
• Stock Photo Platforms: Filter AI-generated uploads, maintain content quality
• Social Media Moderation: Detect AI-generated profile pictures and fake content
• News & Media: Verify photo authenticity, combat misinformation
• NFT Marketplaces: Ensure digital art authenticity
• Academic Research: Study AI image generation patterns

Model Description

This model can:

1. Detect whether an image is real or AI-generated
2. Identify the specific AI generator used (e.g., Stable Diffusion, DALL-E 3, Midjourney, StyleGAN2, etc.)

Architecture

• Base Model: microsoft/beit-large-patch16-224 (BEiT-Large)
• Parameters: ~304M
• Input Size: 224x224 pixels
• Mode: Multi-class classification

Output Labels (33 Classes)

Real: All real images (from ImageNet, COCO, FFHQ, CelebA-HQ, AFHQ, LSUN, MetFaces, Landscape)

AI Sources:

Label	Description
stable_diffusion	Stable Diffusion 1.x/2.x/XL
latent_diffusion	Latent Diffusion Models
glide	OpenAI GLIDE
dalle	DALL-E series
midjourney	Midjourney
stylegan1	StyleGAN v1
stylegan2	StyleGAN v2
stylegan3	StyleGAN v3
pro_gan	Progressive GAN
big_gan	BigGAN
gau_gan	GauGAN / NVIDIA Canvas
cycle_gan	CycleGAN
star_gan	StarGAN
ddpm	Denoising Diffusion Probabilistic Models
vq_diffusion	VQ-Diffusion
palette	Palette Diffusion
gansformer	GANsformer
projected_gan	Projected GAN
diffusion_gan	Diffusion GAN
denoising_diffusion_gan	Denoising Diffusion GAN
taming_transformer	Taming Transformers
generative_inpainting	Generative Inpainting
lama	LaMa Inpainting
mat	MAT Inpainting
cips	CIPS
face_synthetics	Face Synthetics
sfhq	Synthetic Faces HQ

Recommended Usage

from transformers import AutoModelForImageClassification, AutoImageProcessor
from PIL import Image
import torch

# Load model
model_id = "boluobobo/ItsNotAI-ai-detector-v1"
model = AutoModelForImageClassification.from_pretrained(model_id)
processor = AutoImageProcessor.from_pretrained(model_id)

def detect_image(image_path):
    image = Image.open(image_path).convert("RGB")
    inputs = processor(image, return_tensors="pt")

    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=-1)[0]

    # Get top predictions
    top_indices = probs.argsort(descending=True)[:5]
    predictions = []
    for idx in top_indices:
        label = model.config.id2label[str(idx.item())]
        score = probs[idx].item()
        predictions.append({"label": label, "score": round(score, 3)})

    return {
        "predicted_source": predictions[0]["label"],
        "confidence": predictions[0]["score"],
        "top5_predictions": predictions
    }

# Example
result = detect_image("test.jpg")
print(f"Predicted Source: {result['predicted_source']}")
print(f"Confidence: {result['confidence']:.1%}")

Example Output:

{
  "predicted_source": "stable_diffusion",
  "confidence": 0.452,
  "top5_predictions": [
    {"label": "stable_diffusion", "score": 0.452},
    {"label": "latent_diffusion", "score": 0.213},
    {"label": "glide", "score": 0.089},
    {"label": "ddpm", "score": 0.056},
    {"label": "imagenet", "score": 0.042}
  ]
}

Note: v1 returns source classification only. For binary Real/AI detection, use v2 with its dedicated binary head.

Performance

Metric	Value
Accuracy	93.51%
Precision	95.40%
Recall	93.51%
F1 Score	94.11%

Quick Start

Installation

pip install transformers torch pillow

Basic Usage

from transformers import AutoModelForImageClassification, AutoImageProcessor
from PIL import Image
import torch

# Load model
model_id = "boluobobo/ItsNotAI-ai-detector-v1"
model = AutoModelForImageClassification.from_pretrained(model_id)
processor = AutoImageProcessor.from_pretrained(model_id)

# Load image
image = Image.open("your_image.jpg").convert("RGB")

# Predict
inputs = processor(image, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    pred_idx = outputs.logits.argmax(-1).item()

# Get label
label = model.config.id2label[pred_idx]
confidence = probs[0][pred_idx].item()

print(f"Prediction: {label}")
print(f"Confidence: {confidence:.2%}")

With Source Metadata

import json
from huggingface_hub import hf_hub_download

# Download source metadata
meta_path = hf_hub_download(repo_id=model_id, filename="source_meta.json")
with open(meta_path) as f:
    meta = json.load(f)

source_names = meta["source_names"]

# Get all probabilities
for i, (name, prob) in enumerate(zip(source_names, probs[0].tolist())):
    if prob > 0.01:  # Show only >1%
        print(f"  {name}: {prob:.2%}")

Training Details

• Dataset: ArtiFact (50K+ images from multiple sources)
• Epochs: 10
• Batch Size: 64
• Learning Rate: 5e-6
• Optimizer: AdamW with cosine scheduler
• Loss: Focal Loss with label smoothing (0.1)
• Hardware: NVIDIA T4 / A100 GPU

FAQ

Q: Can this detect Midjourney images? A: Yes, the model can detect images from Midjourney, Stable Diffusion, DALL-E, and 25+ other AI generators.

Q: Does it work on digital paintings? A: Yes! We specialize in digital art, manga, and illustration detection with input from 100+ professional artists.

Q: How is this different from other AI detectors? A: ItsNotAI focuses on helping artists prove their work is human-made, not just catching AI usage. We provide verifiable labels for authentic artwork.

Q: What image formats are supported? A: PNG, JPG, WEBP, and other common formats. Images are automatically resized to 224x224 for processing.

Limitations

• Best performance on 224x224 or larger images
• May have reduced accuracy on heavily compressed images
• Trained primarily on Western-style images
• New AI generators not in training data may not be correctly identified

Citation

@misc{itsnotai2025,
  title={ItsNotAI: Multi-class AI Image Detection},
  author={ItsNotAI Team},
  year={2025},
  url={https://huggingface.co/boluobobo/ItsNotAI-ai-detector-v1}
}

License

Apache 2.0