Modnet

🚀 wuchendi/MODNet (Matting Objective Decomposition Network)

实时无Trimap人像抠图模型

本项目是用于人像抠图的MODNet模型，支持使用transformers.js库进行推理，可实现实时无Trimap的人像抠图功能。

• 代码仓库：https://github.com/WuChenDi/MODNet
• 演示空间：https://huggingface.co/spaces/wuchendi/MODNet
• SwanLab演示：https://swanlab.cn/@wudi/MODNet/overview

🚀 快速开始

安装依赖

首先，你需要使用PNPM安装@huggingface/transformers库：

pnpm add @huggingface/transformers

运行示例代码

使用以下代码，结合wuchendi/MODNet模型进行人像抠图：

/* eslint-disable no-console */
import { AutoModel, AutoProcessor, RawImage } from '@huggingface/transformers'

async function main() {
  try {
    console.log('üöÄ Initializing MODNet...')

    // Load model
    console.log('📦 Loading model...')
    const model = await AutoModel.from_pretrained('wuchendi/MODNet', {
      dtype: 'fp32',
      progress_callback: (progress) => {
        // @ts-ignore
        if (progress.progress) {
          // @ts-ignore
          console.log(`Model loading progress: ${(progress.progress).toFixed(2)}%`)
        }
      }
    })
    console.log('‚úÖ Model loaded successfully')

    // Load processor
    console.log('üîß Loading processor...')
    const processor = await AutoProcessor.from_pretrained('wuchendi/MODNet', {})
    console.log('‚úÖ Processor loaded successfully')

    // Load image from URL
    const url = 'https://res.cloudinary.com/dhzm2rp05/image/upload/samples/logo.jpg'
    console.log('🖼️ Loading image:', url)
    const image = await RawImage.fromURL(url)
    console.log('‚úÖ Image loaded successfully', `Dimensions: ${image.width}x${image.height}`)

    // Pre-process image
    console.log('🔄 Preprocessing image...')
    const { pixel_values } = await processor(image)
    console.log('‚úÖ Image preprocessing completed')

    // Generate alpha matte
    console.log('🎯 Generating alpha matte...')
    const startTime = performance.now()
    const { output } = await model({ input: pixel_values })
    const inferenceTime = performance.now() - startTime
    console.log('‚úÖ Alpha matte generated', `Time: ${inferenceTime.toFixed(2)}ms`)

    // Save output mask
    console.log('üíæ Saving output...')
    const mask = await RawImage.fromTensor(output[0].mul(255).to('uint8')).resize(image.width, image.height)
    await mask.save('src/assets/mask.png')
    console.log('‚úÖ Output saved to assets/mask.png')

  } catch (error) {
    console.error('‚ùå Error during processing:', error)
    throw error
  }
}

main().catch(console.error)

示例结果

输入图像	输出掩码

📄 许可证

本项目采用Apache-2.0许可证。