AIRealNet-2

AIRealNet-2

A small image classifier that tells AI-generated pictures from human photographs, tuned for social media images of different quality. It suits moderators, researchers and fact-checkers who screen images.

Made by Modotte on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AIRealNet-2 is a binary image classifier that sorts pictures into AI-generated or human photograph. It is the upgraded version of AIRealNet, built on a SwinV2 Tiny backbone with about 195 million parameters. This version was tuned for social media images that have been compressed in different ways, which fooled the older model. The training set holds 98k images, split evenly between AI and human pictures, and personal photos were left out on purpose.

Who it suits: It suits moderators, researchers, fact-checkers and developers who need a first screen for synthetic images on social media, in publications or on media platforms.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Trust and safety | Content moderators | Flag likely AI-generated images for review | Screen an uploaded picture before a person checks it
News and media | Fact-checkers | Get a first signal on a suspicious image | Run a viral picture through the classifier
Research | Synthetic media researchers | Use a small open detector as a baseline | Test it on images from a new generator

If you need proof for a legal case, you can skip it. The card says it is not meant for legal verification without human corroboration.

Where to use it: The record lists the transformers library, so Python is the main place. The model is small enough for a normal workstation or a modest server.

• Python with the transformers image-classification pipeline
• A moderation or review tool that needs a quick signal
• Windows, macOS or Linux with Python, to download the files

Title: Jobs it does
Use case | Who | What they get back
AI image screening | Moderators | A label and a score for each picture
Synthetic media research | Researchers | A baseline detector to compare against
Fact-checking support | Newsrooms | A first signal to confirm with human review

How to set it up: You need Python with pip and the transformers library, plus memory for about 781 MB of F32 weights. The card gives a short demo you can follow after you install transformers.

• Python with pip
• The transformers library
• Memory for about 781 MB of F32 weights, or about 98 MB at 4-bit by estimate

1. Install the transformers library.
2. Download the model files, or let the pipeline fetch them.
3. Run the pipeline on a picture.

pip install -U transformers

pip install -U huggingface_hub
hf download Modotte/AIRealNet-2

How to use it: Create an image-classification pipeline with the model name, then pass it a picture. It returns two labels with scores, one for a real photograph and one for an artificial image. Class 0 means AI-generated and class 1 means a human photograph. Images are resized to 256 by 256 before they reach the model.

from transformers import pipeline

pipe = pipeline("image-classification", model="Modotte/AIRealNet-2")

Example: The card shows a photo that the older model called fake because of social media compression. AIRealNet-2 labels it as a real photograph with a score close to 1. You still read the score as a hint, not a verdict.

For newcomers and experts: If you are new, think of it as a quick check that guesses whether a picture came from a camera or a generator. Try a few images you already know the source of, to get a feel for how it behaves.

If you know the field, the model resizes the image, extracts features with the SwinV2 Tiny vision transformer and applies a binary head. It was fine-tuned on the train split of about 98k social media images, with validation on a separate balanced subset. The card lists its blind spots: subtle edits, AI-edited pictures of real people, and generators it has not seen.

The image is resized to 256 by 256
SwinV2 Tiny extracts features
A binary head scores the two classes
You get AI-generated or human photograph

Good to know: The training data is balanced between the two classes, which keeps the model from favoring one.

Title: Training data split
AI-generated: 50%
Human images: 50%

• License: MIT
• Cost to run: none named
• Data it reaches: the model files from Hugging Face, and the images you pass in
• Not for: legal verification without human checks, tiny edits, or AI-edited pictures of real people, which it does not detect

On Hugging Face, 13 November 2026

• Task: image classification
• Library: transformers
• Parameters: 195M
• Weights: about 781 MB in F32; about 98 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 492
• Likes: 10
• License: mit
• Based on: Modotte/AIRealNet
• Tags: transformers, safetensors, swinv2, image-classification, safety, modotte, airealnet, sota

Links

• Model: https://huggingface.co/Modotte/AIRealNet-2
• Files: https://huggingface.co/Modotte/AIRealNet-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Modotte/AIRealNet-2

Then follow the usage in its model card. Nothing in this kit runs by itself.