|
|
--- |
|
|
license: apache-2.0 |
|
|
datasets: |
|
|
- nebula/OpenSDI_test |
|
|
- madebyollin/megalith-10m |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- google/siglip2-base-patch16-224 |
|
|
pipeline_tag: image-classification |
|
|
library_name: transformers |
|
|
tags: |
|
|
- OpenSDI |
|
|
- Spotting Diffusion-Generated Images in the Open World |
|
|
- OpenSDI |
|
|
- SD3 |
|
|
- AI-vs-Real |
|
|
- SigLIP2 |
|
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
# OpenSDI-SD3-SigLIP2 |
|
|
|
|
|
> OpenSDI-SD3-SigLIP2 is a vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for binary image classification. It is trained to detect whether an image is a real photograph or generated using Stable Diffusion 3 (SD3), using the SiglipForImageClassification architecture. |
|
|
|
|
|
> [!note] |
|
|
*SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features* https://arxiv.org/pdf/2502.14786 |
|
|
|
|
|
> [!note] |
|
|
*OpenSDI: Spotting Diffusion-Generated Images in the Open World* https://arxiv.org/pdf/2503.19653, OpenSDI SD3 SigLIP2 works best with crisp and high-quality images. Noisy images are not recommended for validation. |
|
|
|
|
|
> [!warning] |
|
|
If the task is based on image content moderation or AI-generated image vs. real image classification, it is recommended to use the OpenSDI-Flux.1-SigLIP2 model. |
|
|
|
|
|
```py |
|
|
Classification Report: |
|
|
precision recall f1-score support |
|
|
|
|
|
Real_Image 0.8526 0.8916 0.8716 10000 |
|
|
SD3_Generated 0.8864 0.8458 0.8656 10000 |
|
|
|
|
|
accuracy 0.8687 20000 |
|
|
macro avg 0.8695 0.8687 0.8686 20000 |
|
|
weighted avg 0.8695 0.8687 0.8686 20000 |
|
|
``` |
|
|
|
|
|
 |
|
|
|
|
|
--- |
|
|
|
|
|
## Label Space: 2 Classes |
|
|
|
|
|
The model classifies an image as either: |
|
|
|
|
|
``` |
|
|
Class 0: Real_Image |
|
|
Class 1: SD3_Generated |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Install Dependencies |
|
|
|
|
|
```bash |
|
|
pip install -q transformers torch pillow gradio hf_xet |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Inference Code |
|
|
|
|
|
```python |
|
|
import gradio as gr |
|
|
from transformers import AutoImageProcessor, SiglipForImageClassification |
|
|
from PIL import Image |
|
|
import torch |
|
|
|
|
|
# Load model and processor |
|
|
model_name = "prithivMLmods/OpenSDI-SD3-SigLIP2" # Update with the correct model path |
|
|
model = SiglipForImageClassification.from_pretrained(model_name) |
|
|
processor = AutoImageProcessor.from_pretrained(model_name) |
|
|
|
|
|
# Label mapping |
|
|
id2label = { |
|
|
"0": "Real_Image", |
|
|
"1": "SD3_Generated" |
|
|
} |
|
|
|
|
|
def classify_image(image): |
|
|
image = Image.fromarray(image).convert("RGB") |
|
|
inputs = processor(images=image, return_tensors="pt") |
|
|
|
|
|
with torch.no_grad(): |
|
|
outputs = model(**inputs) |
|
|
logits = outputs.logits |
|
|
probs = torch.nn.functional.softmax(logits, dim=1).squeeze().tolist() |
|
|
|
|
|
prediction = { |
|
|
id2label[str(i)]: round(probs[i], 3) for i in range(len(probs)) |
|
|
} |
|
|
|
|
|
return prediction |
|
|
|
|
|
# Gradio Interface |
|
|
iface = gr.Interface( |
|
|
fn=classify_image, |
|
|
inputs=gr.Image(type="numpy"), |
|
|
outputs=gr.Label(num_top_classes=2, label="SD3 Image Detection"), |
|
|
title="OpenSDI-SD3-SigLIP2", |
|
|
description="Upload an image to determine whether it is a real photograph or generated by Stable Diffusion 3 (SD3)." |
|
|
) |
|
|
|
|
|
if __name__ == "__main__": |
|
|
iface.launch() |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
OpenSDI-SD3-SigLIP2 is designed for tasks such as: |
|
|
|
|
|
* Generative Image Analysis – Identify SD3-generated images for benchmarking and quality inspection. |
|
|
* Dataset Validation – Ensure training or evaluation datasets are free from unintended generative artifacts. |
|
|
* Content Authenticity – Verify whether visual media originates from real-world photography or AI generation. |
|
|
* Digital Forensics – Aid in determining the origin of visual content in investigative scenarios. |
