A Highly Efficient Model for NSFW Detection. Very effective for pre-publication image and video control, or for limiting children's access to harmful publications. You can either just predict the classes and their boundingboxes or even mask the predicted harmful object(s) or mask the entire image. Please see the deployment codes below.

Developed by:
- Lê Chí Tài (tai.le@erax.ai)
- Phạm Đình Thục (thuc.pd@erax.ai)
- Mr. Nguyễn Anh Nguyên (nguyen@erax.ai)
Model version: v1.1
License: Apache 2.0

Model Details / Overview

Model Architecture: YOLO11 (Nano, Small, Medium)
Task: Object Detection (NSFW Detection)
Dataset: Private datasets (From Internet).
Training set: 40192 images.
Validation set: 3495 images.
Classes: anus, make_love, nipple, penis, vagina.

Labels

Training Configuration

Model Weights Files:
- Nano: erax-anti-nsfw-yolo11n-v1.1.pt (5.45 MB)
- Small: erax-anti-nsfw-yolo11s-v1.1.pt (40.5 MB)
- Medium: erax-anti-nsfw-yolo11m-v1.1.pt (19.2 MB)
Training Config:
- Number of Epochs: 100
- Learning Rate: 0.01
- Batch Size: 336/192/92 (Nano/Small/Medium)
- Image Size: 640x640
- Training server: 4 x NVIDIA RTX A4000 (16GB GDDR6)

Evaluation Metrics

Below are the key metrics from the model evaluation on the validation set: comming soon

Benchmark

CPU: 11th Gen Intel Core(TM) i7-11800H 2.30GHz
GPU: NVIDIA GeForce RTX 3050 Ti 3902MiB

Format	Model	Metrics/mAP50-95(B)	GPU	CPU
Inference time (ms/im)	FPS	Inference time (ms/im)	FPS
PyTorch	erax-anti-nsfw-yolo11n-v1.1.pt	0.438	3.500	286	27.900	36
erax-anti-nsfw-yolo11s-v1.1.pt	0.453	7.000	143	71.000	14
erax-anti-nsfw-yolo11m-v1.1.pt	0.467	16.500	61	206.600	5
TorchScript	erax-anti-nsfw-yolo11n-v1.1.torchscript	0.435	3.700	270	38.500	26
erax-anti-nsfw-yolo11s-v1.1.torchscript	0.449	8.100	123	108.500	9
erax-anti-nsfw-yolo11m-v1.1.torchscript	0.463	20.300	49	394.900	3
ONNX	erax-anti-nsfw-yolo11n-v1.1.onnx	0.435	-	-	28.300	35
erax-anti-nsfw-yolo11s-v1.1.onnx	0.449	-	-	59.800	17
erax-anti-nsfw-yolo11m-v1.1.onnx	0.463	-	-	157.800	6
OpenVINO	erax-anti-nsfw-yolo11n-v1.1_openvino_model	0.435	13.900	72	15.900	63
erax-anti-nsfw-yolo11s-v1.1_openvino_model	0.449	72.300	14	40.800	25
erax-anti-nsfw-yolo11m-v1.1_openvino_model	0.463	245.900	4	121.700	8
TensorRT	erax-anti-nsfw-yolo11n-v1.1.engine	0.435	3.500	286	-	-
erax-anti-nsfw-yolo11s-v1.1.engine	0.449	6.800	147	-	-
erax-anti-nsfw-yolo11m-v1.1.engine	0.463	15.700	64	-	-
PaddlePaddle	erax-anti-nsfw-yolo11n-v1.1_paddle_model	0.435	214.700	5	136.200	7
erax-anti-nsfw-yolo11s-v1.1_paddle_model	0.449	517.700	2	234.600	4
erax-anti-nsfw-yolo11m-v1.1_paddle_model	0.463	887.000	1	506.300	2
MNN	erax-anti-nsfw-yolo11n-v1.1.mnn	0.435	55.800	18	59.300	17
erax-anti-nsfw-yolo11s-v1.1.mnn	0.449	147.600	7	146.300	7
erax-anti-nsfw-yolo11m-v1.1.mnn	0.463	378.500	3	380.700	3
NCNN	erax-anti-nsfw-yolo11n-v1.1_ncnn_model	0.435	57.100	18	61.100	16
erax-anti-nsfw-yolo11s-v1.1_ncnn_model	0.449	141.200	7	137.200	7
erax-anti-nsfw-yolo11m-v1.1_ncnn_model	0.463	375.500	3	367.400	3

Training Validation Results

Training and Validation Losses

Confusion Matrix

Inference

To use the trained model, follow these steps:

Install the necessary packages:

pip install ultralytics supervision huggingface-hub

Download Pretrained model:

python

from huggingface_hub import snapshot_download
snapshot_download(repo_id="erax-ai/EraX-Anti-NSFW-V1.1", local_dir="./", force_download=True)

Simple Use Case:

python

from ultralytics import YOLO
from PIL import Image
import supervision as sv
import numpy as np

IOU_THRESHOLD        = 0.3
CONFIDENCE_THRESHOLD = 0.2

# pretrained_path = "erax-anti-nsfw-yolo11m-v1.1.pt"
# pretrained_path = "erax-anti-nsfw-yolo11s-v1.1.pt"
pretrained_path = "erax-anti-nsfw-yolo11n-v1.1.pt"

image_path_list = ["test_images/img_1.jpg", "test_images/img_2.jpg"]

model = YOLO(pretrained_path)
results = model(image_path_list,
                  conf=CONFIDENCE_THRESHOLD,
                  iou=IOU_THRESHOLD
                )


for result in results:
    annotated_image = result.orig_img.copy()
    h, w = annotated_image.shape[:2]
    anchor = h if h > w else w

    detections = sv.Detections.from_ultralytics(result)
    label_annotator = sv.LabelAnnotator(text_color=sv.Color.BLACK,
                                        text_position=sv.Position.CENTER,
                                        text_scale=anchor/1700)
    
    pixelate_annotator = sv.PixelateAnnotator(pixel_size=anchor/50)
    
    annotated_image = pixelate_annotator.annotate(
        scene=annotated_image.copy(),
        detections=detections
    )


    annotated_image = label_annotator.annotate(
        annotated_image,
        detections=detections
    )

    
    sv.plot_image(annotated_image, size=(10, 10))

More examples

Example 01:
Example 03
Example 02:
Example 06
Example 03: SAFEEST for using make_love class as it will cover entire context. Without make_love class | With make_love class :-------------------------:|:-------------------------:
|

|

|

Citation

If you find our project useful, we would appreciate it if you could star our repository and cite our work as follows:

bibtex

@article{EraX-Anti-NSFW-V1.1,
  author    = {Lê Chí Tài and
              Phạm Đình Thục and
              Mr. Nguyễn Anh Nguyên and
              Đoàn Thành Khang and
              Mr. Trần Hải Khương and
              Mr. Trương Công Đức and 
              Phan Nguyễn Tuấn Kha and 
              Phạm Huỳnh Nhật},
  title     = {EraX-Anti-NSFW-V1.1: A Highly Efficient Model for NSFW Detection},
  organization={EraX JS Company},
  year={2024},
  url={https://huggingface.co/erax-ai/EraX-Anti-NSFW-V1.1}
}