@avatune/skin-tone-predictor

Source: packages/skin-tone-predictor/README.md

Browser-based skin tone prediction using TensorFlow.js. Classifies skin tone into 3 categories: dark, medium, light.

Tiny model (~110KB) with blazingly fast loading and inference in the browser.

Installation

npm

npm install @avatune/skin-tone-predictor @tensorflow/tfjs

yarn

yarn add @avatune/skin-tone-predictor @tensorflow/tfjs

pnpm

pnpm add @avatune/skin-tone-predictor @tensorflow/tfjs

bun

bun add @avatune/skin-tone-predictor @tensorflow/tfjs

Usage

import { createSkinTonePredictor } from '@avatune/skin-tone-predictor'

// Uses jsDelivr CDN by default - no setup required!
const predictor = createSkinTonePredictor()
await predictor.loadModel()

const result = await predictor.predictFromImage(imageElement)
console.log(result)
// {
//   tone: 'medium',
//   confidence: 0.84,
//   probabilities: { dark: 0.08, medium: 0.84, light: 0.06 },
//   faceDetected: true
// }

Model Files

By default, models are loaded from jsDelivr CDN (https://cdn.jsdelivr.net/npm/@avatune/skin-tone-predictor@1.2.2/dist/model). No setup required!

Self-hosting (Optional)

If you prefer to self-host the model files, copy them from dist/model/ to your public directory:

• model.json - Model architecture and weights manifest
• classes.json - Class labels
• group1-shard1of1.bin - Model weights

Then pass the path to the predictor:

const predictor = createSkinTonePredictor('/models/skin-tone')

Setup with Vite (Optional)

import { copyFileSync, mkdirSync, readdirSync } from 'node:fs'
import { join } from 'node:path'
import { defineConfig } from 'vite'

export default defineConfig({
  plugins: [
    {
      name: 'copy-tfjs-models',
      buildStart() {
        const srcDir = join(__dirname, 'node_modules', '@avatune', 'skin-tone-predictor', 'dist', 'model')
        const destDir = join(__dirname, 'public', 'models', 'skin-tone')

        mkdirSync(destDir, { recursive: true })

        copyFileSync(join(srcDir, 'model.json'), join(destDir, 'model.json'))
        copyFileSync(join(srcDir, 'classes.json'), join(destDir, 'classes.json'))

        const files = readdirSync(srcDir)
        for (const file of files) {
          if (file.endsWith('.bin')) {
            copyFileSync(join(srcDir, file), join(destDir, file))
          }
        }

        console.log('✓ Copied skin-tone model to public/models')
      },
    },
  ],
})

API

Constructor

createSkinTonePredictor(modelDir?: string)

Parameters:

• modelDir (optional) - Path to directory containing model files. Defaults to jsDelivr CDN

Methods

loadModel(): Promise<void>

Loads the TFJS model and class labels. Call this once before making predictions.

predict(imageTensor: tf.Tensor3D): Promise<SkinToneResult>

Predicts skin tone from an image tensor.

Parameters:

• imageTensor - Normalized RGB image tensor [H, W, 3] with values in range [0, 1]

Returns:

{
  tone: string         // Predicted class: 'dark' | 'medium' | 'light'
  confidence: number   // Confidence score [0, 1]
  probabilities: Record<string, number>  // Scores for all classes
}

Model Details

• Architecture: MobileNetV2-based CNN
• Input: 128x128 RGB images
• Classes: 3 (dark, medium, light)
• Training: FairFace dataset
• Accuracy: ~68%
• Model size: ~110KB (uint8 quantized)
• Format: TensorFlow.js with uint8 quantization

License

See LICENSE.md for license information.