Open-weights image classifier for content moderation: safe, suggestive or explicit.
Model on Hugging Face
Blue-Eye predicts whether an image is safe, suggestive or explicit, with a probability for each class. It is a DINOv3 ViT-L/16 fine-tuned end to end at 512x512 on about 2.6 million images, and it handles both real photographs and anime/illustration.
On a benchmark of 3,000 real photographs, Blue-Eye reaches 88.9% accuracy, ahead of AWS Rekognition (87.3%), Gemini 3.1 Pro (86.4%), Google Cloud Vision (83.5%) and every open-source NSFW detector it was compared with. All systems were evaluated on the same images.
The weights (1.2 GB) are hosted on Hugging Face; this repository contains the inference code. Full evaluation details are in the model card.
git clone https://github.com/prnshu-p/Blue-Eye.git
cd Blue-Eye
pip install -r requirements.txt huggingface_hub
The weights are gated: open the model page, click
"Agree and download" once, then log in with hf auth login.
from huggingface_hub import snapshot_download
from inference import classify
path = snapshot_download("pp1618/Blue-Eye")
for result in classify(["photo.jpg", "drawing.png"], model=path):
print(result["label"], result["probabilities"])
From the command line:
python inference.py photo.jpg folder_of_images/ --model pp1618/Blue-Eye --device cuda
Add --precision bf16 on GPUs with bfloat16 support for faster inference.
| Class | Covers |
|---|---|
safe | no sexualised content, including swimwear, fitness, medical images and non-sexual art |
suggestive | sexualised but not explicit, such as posed lingerie shots |
explicit | exposed genitalia, sexual acts or full nudity |
Blue-Eye is built for moderating sexual content: filtering feeds and search results, blurring images, age-gating adult content, prioritising images for human review, and curating datasets. It covers sexual content only and is not intended for age estimation, child-safety detection, or monitoring individual people.
Built on Meta's DINOv3 and released under the
DINOv3 License (see LICENSE).
Commercial use is permitted under its terms.
@misc{patel2026blueeye,
title = {Blue-Eye: a content-safety image classifier},
author = {Patel, Pranshu},
year = {2026},
url = {https://huggingface.co/pp1618/Blue-Eye}
}
Open-weights image classifier for content moderation: safe, suggestive or explicit.
Model on Hugging Face
Blue-Eye predicts whether an image is safe, suggestive or explicit, with a probability for each class. It is a DINOv3 ViT-L/16 fine-tuned end to end at 512x512 on about 2.6 million images, and it handles both real photographs and anime/illustration.
On a benchmark of 3,000 real photographs, Blue-Eye reaches 88.9% accuracy, ahead of AWS Rekognition (87.3%), Gemini 3.1 Pro (86.4%), Google Cloud Vision (83.5%) and every open-source NSFW detector it was compared with. All systems were evaluated on the same images.
The weights (1.2 GB) are hosted on Hugging Face; this repository contains the inference code. Full evaluation details are in the model card.
git clone https://github.com/prnshu-p/Blue-Eye.git
cd Blue-Eye
pip install -r requirements.txt huggingface_hub
The weights are gated: open the model page, click
"Agree and download" once, then log in with hf auth login.
from huggingface_hub import snapshot_download
from inference import classify
path = snapshot_download("pp1618/Blue-Eye")
for result in classify(["photo.jpg", "drawing.png"], model=path):
print(result["label"], result["probabilities"])
From the command line:
python inference.py photo.jpg folder_of_images/ --model pp1618/Blue-Eye --device cuda
Add --precision bf16 on GPUs with bfloat16 support for faster inference.
| Class | Covers |
|---|---|
safe | no sexualised content, including swimwear, fitness, medical images and non-sexual art |
suggestive | sexualised but not explicit, such as posed lingerie shots |
explicit | exposed genitalia, sexual acts or full nudity |
Blue-Eye is built for moderating sexual content: filtering feeds and search results, blurring images, age-gating adult content, prioritising images for human review, and curating datasets. It covers sexual content only and is not intended for age estimation, child-safety detection, or monitoring individual people.
Built on Meta's DINOv3 and released under the
DINOv3 License (see LICENSE).
Commercial use is permitted under its terms.
@misc{patel2026blueeye,
title = {Blue-Eye: a content-safety image classifier},
author = {Patel, Pranshu},
year = {2026},
url = {https://huggingface.co/pp1618/Blue-Eye}
}