Generate image captions using Hugging Face | Supabase 简体中文

We can combine Hugging Face with Supabase Storage and Database Webhooks to automatically caption for any image we upload to a storage bucket.

About Hugging Face

Hugging Face is the collaboration platform for the machine learning community.

Huggingface.js provides a convenient way to make calls to 100,000+ Machine Learning models, making it easy to incorporate AI functionality into your Supabase Edge Functions.

Setup

Open your Supabase project dashboard or create a new project.
Create a new bucket called images.
Generate TypeScript types from remote Database.
Create a new Database table called image_caption.
- Create id column of type uuid which references storage.objects.id.
- Create a caption column of type text.
Regenerate TypeScript types to include new image_caption table.
Deploy the function to Supabase: supabase functions deploy huggingface-image-captioning.
Create the Database Webhook in the Supabase Dashboard to trigger the huggingface-image-captioning function anytime a record is added to the storage.objects table.

Generate TypeScript types

To generate the types.ts file for the storage and public schemas, run the following command in the terminal:

supabase gen types typescript --project-id=your-project-ref --schema=storage,public > supabase/functions/huggingface-image-captioning/types.ts

Code

Find the complete code on GitHub.

import { HfInference } from 'https://esm.sh/@huggingface/inference@2.3.2'
import { createClient } from 'npm:@supabase/supabase-js@2'

import { Database } from './types.ts'

console.log('Hello from `huggingface-image-captioning` function!')

const hf = new HfInference(Deno.env.get('HUGGINGFACE_ACCESS_TOKEN'))

type SoRecord = Database['storage']['Tables']['objects']['Row']
interface WebhookPayload {
  type: 'INSERT' | 'UPDATE' | 'DELETE'
  table: string
  record: SoRecord
  schema: 'public'
  old_record: null | SoRecord
}

Deno.serve(async (req) => {
  const payload: WebhookPayload = await req.json()
  const soRecord = payload.record
  const SUPABASE_SECRET_KEYS = JSON.parse(Deno.env.get('SUPABASE_SECRET_KEYS')!)
  const supabaseAdminClient = createClient<Database>(
    // Supabase API URL - env var exported by default when deployed.
    Deno.env.get('SUPABASE_URL') ?? '',
    // Supabase API SECRET KEY - env var exported by default when deployed.
    SUPABASE_SECRET_KEYS['default'] ?? ''
  )

  // Construct image url from storage
  const { data, error } = await supabaseAdminClient.storage
    .from(soRecord.bucket_id!)
    .createSignedUrl(soRecord.path_tokens!.join('/'), 60)
  if (error) throw error
  const { signedUrl } = data

  // Run image captioning with Huggingface
  const imgDesc = await hf.imageToText({
    data: await (await fetch(signedUrl)).blob(),
    model: 'nlpconnect/vit-gpt2-image-captioning',
  })

  // Store image caption in Database table
  await supabaseAdminClient
    .from('image_caption')
    .insert({ id: soRecord.id!, caption: imgDesc.generated_text })
    .throwOnError()

  return new Response('ok')
})