ygexe / gpupixel

Cross-Platform AI Beauty Effects Library, Achieving Commercial-Grade Beauty Effects. Written in C++11, Based on OpenGL/ES and VNN.

GPUPixel

---

English | 简体中文

GPUPixel @ PixPark

Introduction

⛰️ GPUPixel is a high-performance image and video AI beauty effect library written in C++11, Extremely easy to compile and integrate, with a very small library size.

🔑 It is GPU-based and comes with built-in beauty effects filters that can achieve commercial-grade results.

🔌 It supports platforms including iOS, Mac, Android, and it can theoretically be ported to any platform that supports OpenGL/ES.

💡 The face key points detection currently utilizes the Face++ library, but it will be replaced with either VNN or NCNN in the future.

Effects Preview

👉 Video: YouTube | BiliBili

Origin	Smooth	White	ThinFace
BigEye	Lipstick	Blusher	ON-OFF

Architecture

Features Compared

This table compares the features supported by GPUPixel and GPUImage:

✅: Supported | ❌: Not supported | ✏️: Planning

	GPUPixel	GPUImage
🍎Filters:	✅	❌
Skin Smoothing Filter	✅	❌
Skin Whitening Filter	✅	❌
Face Slimming Filter	✅	❌
Big Eyes Filter	✅	❌
Lipstick Filter	✅	❌
Blush Filter	✅	❌
More Build in Filter	✅	✅
🍓Input Formats:
YUV420P(I420)	✅	❌
RGBA	✅	✅
JPEG	✅	✅
PNG	✅	✅
NV21(for Android)	✏️	❌
🍉Output Formats:
RGBA	✅	✅
YUV420P(I420)	✏️	❌
🥑Platform:
iOS	✅	✅
Mac	✅	✅
Android	✅	❌
Win	✏️	❌
Linux	✏️	❌

System Requirements

OS	iOS	OSX	Android	Windows	Linux
Min Support OS Version	10.0	10.13	5.0 (API 21)	-	-

Performance

iPhone

-	iPhone 6P	iPhone 8	iPhone X	iPhone 11	iPhone 14 pro
CPU	5%	5%	3%	3%	3%
Time Taken	10ms	4ms	3ms	3ms	3ms

Android

-	Xiaomi 10	Huawei Meta30	Vivo	SAMSUNG	Google Pixel
CPU	3%	5%	-	-	-
Time Taken	6ms	5ms	-	-	-

Compilation

iOS

1. Open objc/gpupixel.xcodeproj or objc/demo/PixDemo.xcodeproj Xcode project ;

2. Select the corresponding platform for compilation;

Android

1. Open the directory ./android in Android Studio

2. Configure NDK r21+

Library Size

	iOS(.framework)	MacOS(.framework)	Android(.aar)
Size	2.4 MB	2.6 MB	2.1 MB

Usage

Refer to ./objc/demo or ./android demo

.h file

// video data input
std::shared_ptr<SourceRawDataInput> source_raw_input_;
// beauty filter
std::shared_ptr<FaceBeautyFilter> face_beauty_filter_;
// video data output 
std::shared_ptr<TargetRawDataOutput> target_raw_output_;

Create Filter, .cpp file

 GPUPixel::GPUPixelContext::getInstance()->runSync([&] {
    // Create filter
    source_raw_input_ = SourceRawDataInput::create();
    target_raw_output_ = TargetRawDataOutput::create();
    // Face Beauty Filter
    face_beauty_filter_ = FaceBeautyFilter::create();
    
    // Add filter
    source_raw_input_->addTarget(face_beauty_filter_)->addTarget(target_raw_output_);
 }

Input Image Data I420 or RGBA

// ...
// YUVI420
 source_raw_input_->uploadBytes(width,
                                height, 
                                bufferY,
                                strideY, 
                                bufferU, 
                                strideU,
                                bufferV, 
                                strideV);
// ...
// bytes: RGBA data
 source_raw_input_->uploadBytes(bytes,
                                width, 
                                height, 
                                stride);

Output Data Callback

// I420 callback
target_raw_output_->setI420Callbck([=](const uint8_t *data, 
                                        int width, 
                                        int height, 
                                        int64_t ts) {
    size_t y_size = width * height;
    const uint8_t *uData = data + y_size;
    const uint8_t *vData = data + y_size + y_size / 4;
    // Do something you want
});

// RGBA callback->
target_raw_output_->setPixelsCallbck([=](const uint8_t *data, 
                                        int width, 
                                        int height, 
                                        int64_t ts) {
    size_t rgba_size = width * height*4;
    // Do something you want
});

// Output data callbck

Star History

Acknowledgement

Reference Project

1. GPUImage
2. CainCamera
3. AwemeLike
4. VNN