BlurGuard models

These are the on-device models used by BlurGuard, an Android app that covers nudity and immodest content on screen as it appears. The app downloads a file only when it is needed and checks it against the size and SHA-256 listed below. It then runs the model on the phone's GPU with LiteRT (TensorFlow Lite), falling back to the CPU. Screen images never leave the phone.

Every file is a format conversion of a published model. Nothing was retrained or fine-tuned. Each file keeps the license of the model it came from (see LICENSE.md).

Files

File	Made from	License	Size
`nudenet_320n_640_fp16.tflite`	NudeNet 320n (YOLOv8n), `SimonJoz/nudenet` `320n.pt`	AGPL-3.0	6.2 MB
`nudenet_640m_640_fp16.tflite`	NudeNet 640m (YOLOv8m), `SimonJoz/nudenet` `640m.pt`	AGPL-3.0	51.9 MB
`anime_censor_n_640_fp16.tflite`	`deepghs/anime_censor_detection` `censor_detect_v1.0_n/model.pt`	MIT	6.1 MB
`anime_censor_s_640_fp16.tflite`	`deepghs/anime_censor_detection` `censor_detect_v1.0_s/model.pt`	MIT	22.4 MB
`selfie_multiclass_256x256.tflite`	MediaPipe selfie multiclass segmenter (Google), unchanged	Apache-2.0	16.4 MB

Exact sizes and SHA-256 hashes are in SHA256SUMS.

Detectors (the four YOLOv8 files)

Input: [1, 640, 640, 3] float32, RGB scaled to 0..1, NHWC. Fit the picture inside the square without stretching it, and fill the rest with grey (114, 114, 114), as Ultralytics does.
Output: [1, 4 + classes, 8400] float32. For each of the 8400 anchors there are four box values (centre x, centre y, width, height, in pixels of the 640 square) and then one score per class (0..1, sigmoid already applied). There is no built-in NMS: keep the boxes above your threshold, then run non-maximum suppression.
Shapes are fixed. The LiteRT GPU delegate rejects models with dynamic tensors, so every tensor has a fixed size. Weights are float16; the graph runs in float32 on the CPU and float16 on the GPU.

NudeNet classes, in output order:

0 FEMALE_GENITALIA_COVERED   6 ANUS_EXPOSED       12 FACE_MALE
1 FACE_FEMALE                7 FEET_EXPOSED       13 BELLY_EXPOSED
2 BUTTOCKS_EXPOSED           8 BELLY_COVERED      14 MALE_GENITALIA_EXPOSED
3 FEMALE_BREAST_EXPOSED      9 FEET_COVERED       15 ANUS_COVERED
4 FEMALE_GENITALIA_EXPOSED  10 ARMPITS_COVERED    16 FEMALE_BREAST_COVERED
5 MALE_BREAST_EXPOSED       11 ARMPITS_EXPOSED    17 BUTTOCKS_COVERED

Anime censor classes, in output order: 0 nipple_f, 1 penis, 2 pussy. These two models are made for anime, cartoons and illustrations, not photos.

Skin segmenter

selfie_multiclass_256x256.tflite is Google's MediaPipe model, unchanged. It is copied here so that all the app's downloads come from one pinned place. Original: https://storage.googleapis.com/mediapipe-models/image_segmenter/selfie_multiclass_256x256/float32/latest/selfie_multiclass_256x256.tflite.

Input: [1, 256, 256, 3] float32, RGB scaled to -1..1.
Output: [1, 256, 256, 6], a score per pixel for: background, hair, body skin, face skin, clothes, other.

BlurGuard uses the body-skin class to cover bare arms, legs and shoulders while leaving the face visible.

How the detectors were converted

convert.py shows each step:

Ultralytics export(format="onnx", imgsz=640, dynamic=False, simplify=True, opset=17).
onnx2tf -tb tf_converter -b 1. The TensorFlow Lite converter backend writes fixed shapes.
The float16 output (*_float16.tflite) was kept.
Check: run each TFLite file on the LiteRT CPU interpreter and the ONNX file on ONNX Runtime, using the same letterboxed test pictures. Compare the class scores.

File	Largest score difference against ONNX
`nudenet_320n_640_fp16.tflite`	0.0029
`nudenet_640m_640_fp16.tflite`	0.0009
`anime_censor_n_640_fp16.tflite`	0.0008
`anime_censor_s_640_fp16.tflite`	0.0011

Tools: Ultralytics 8.4.182, onnx 1.20.1, onnx2tf 2.6.9, TensorFlow 2.21.0, ai-edge-litert 2.1.2, ONNX Runtime 1.26.0.

Intended use and limits

These models are for on-device content filtering, such as parental controls. They make mistakes. They miss some content, especially when it is small, blurry, partly hidden or drawn. They also cover some harmless content. NudeNet's face classes guess gender from appearance, so they will be wrong for some people. Do not use them for any decision about a person.

Licenses

The NudeNet files are derived from AGPL-3.0 weights and stay under AGPL-3.0 (licenses/AGPL-3.0.txt). The only change is the format conversion described above.
The anime censor files are derived from MIT-licensed weights and stay under MIT (licenses/MIT-anime_censor_detection.txt).
The MediaPipe file is under Apache-2.0 (licenses/Apache-2.0.txt).

Downloads last month: -