AI & Vectors

Hugging Face Inference API

Hugging Face is an open source hub for AI/ML models and tools. With over 100,000 machine learning models available, Hugging Face provides a great way to integrate specialized AI & ML tasks into your application.

There are 3 ways to use Hugging Face models in your application:

  1. Use the Transformers Python library to perform inference in a Python backend.
  2. Generate embeddings directly in Edge Functions using Transformers.js.
  3. Use Hugging Face's hosted Inference API to execute AI tasks remotely on Hugging Face servers. This guide will walk you through this approach.

AI tasks

Below are some of the types of tasks you can perform with Hugging Face:

Natural language

Computer vision

Audio

See a full list of tasks.

Access token

First generate a Hugging Face access token for your app:

https://huggingface.co/settings/tokens

Name your token based on the app its being used for and the environment. For example, if you are building an image generation app you might create 2 tokens:

  • "My Image Generator (Dev)"
  • "My Image Generator (Prod)"

Since we will be using this token for the inference API, choose the read role.

Edge Functions

Edge Functions are server-side TypeScript functions that run on-demand. Since Edge Functions run on a server, you can safely give them access to your Hugging Face access token.

To create a new Edge Function, navigate to your local project and initialize Supabase if you haven't already:


_10
supabase init

Then create an Edge Function:


_10
supabase functions new text-to-image

Create a file called .env.local to store your Hugging Face access token:


_10
HUGGING_FACE_ACCESS_TOKEN=<your-token-here>

Let's modify the Edge Function to import Hugging Face's inference client and perform a text-to-image request:


_20
import { serve } from 'https://deno.land/[email protected]/http/server.ts'
_20
import { HfInference } from 'https://esm.sh/@huggingface/[email protected]'
_20
_20
const hf = new HfInference(Deno.env.get('HUGGING_FACE_ACCESS_TOKEN'))
_20
_20
serve(async (req) => {
_20
const { prompt } = await req.json()
_20
_20
const image = await hf.textToImage(
_20
{
_20
inputs: prompt,
_20
model: 'stabilityai/stable-diffusion-2',
_20
},
_20
{
_20
use_cache: false,
_20
}
_20
)
_20
_20
return new Response(image)
_20
})

  1. This function creates a new instance of HfInference using the HUGGING_FACE_ACCESS_TOKEN environment variable.

  2. It expects a POST request that includes a JSON request body. The JSON body should include a parameter called prompt that represents the text-to-image prompt that we will pass to Hugging Face's inference API.

  3. Next we call textToImage(), passing in the user's prompt along with the model that we would like to use for the image generation. Today Hugging Face recommends stabilityai/stable-diffusion-2, but you can change this to any other text-to-image model. You can see a list of which models are supported for each task by navigating to their models page and filtering by task.

  4. We set use_cache to false so that repeat queries with the same prompt will produce new images. If the task and model you are using is deterministic (will always produce the same result based on the same input), consider setting use_cache to true for faster responses.

  5. The image result returned from the API will be a Blob. We can pass the Blob directly into a new Response() which will automatically set the content type and body of the response from the image.

Finally let's serve the Edge Function locally to test it:


_10
supabase functions serve --env-file .env.local --no-verify-jwt

Remember to pass in the .env.local file using the --env-file parameter so that the Edge Function can access the HUGGING_FACE_ACCESS_TOKEN.

At this point, you can make an API request to your Edge Function using your preferred frontend framework (Next.js, React, Expo, etc). We can also test from the terminal using curl:


_10
curl --output result.jpg --location --request POST 'http://localhost:54321/functions/v1/text-to-image' \
_10
--header 'Content-Type: application/json' \
_10
--data '{"prompt":"Llama wearing sunglasses"}'

In this example, your generated image will save to result.jpg:

Llama wearing sunglasses example

Next steps

You can now create an Edge Function that invokes a Hugging Face task using your model of choice.

Try running some other AI tasks.

Resources