Skip to main content

Image Generation Using Stable Diffusion

fal-serverless is a serverless platform that enables you to run Python functions on cloud infrastructure. In this example, we will demonstrate how to use fal-serverless for image generation with help from the stable diffusion model.

Step 1: Install fal-serverless and authenticate

pip install fal-serverless
fal-serverless auth login

More info on authentication.

Step 2: Import required libraries

First, we need to import the necessary libraries and define the requirements for the @isolated decorator:

from fal_serverless import isolated
import io

requirements = [

Step 3: Define the generate function

Next, we will define the generate function, which will be responsible for generating the image using the stable diffusion model:

@isolated(requirements=requirements, machine_type="GPU-T4")
def generate(prompt: str):
import torch
import os
import base64
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

model_id = "runwayml/stable-diffusion-v1-5"
os.environ['TRANSFORMERS_CACHE'] = '/data/hugging_face_cache'

pipe = StableDiffusionPipeline.from_pretrained(
pipe ="cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

generator = torch.Generator("cuda")
image = pipe(prompt, generator=generator).images[0]

buf = io.BytesIO(), format="PNG")
return buf

We set the package and machine type requirements as isolated argument. generate function then imports the necessary modules and initializes the stable diffusion pipeline using the specified model ID. It also sets up the Hugging Face transformers cache directory to use fal-serverless persistent /data directory. The pipeline is then moved to the GPU and configured with a custom scheduler. Finally, the function generates the image based on the given prompt and returns it as a byte buffer.

If you are running multiple stable diffusion function calls consecutively, it may be beneficial to consider using @cached functions to optimize performance for recurring tasks.

Step 4: Generate the image

Now that we have defined the generate function, we can use it to generate an image by passing a prompt:

image_data = generate("Donkey walking on clouds")

This will generate an image based on the given prompt "Donkey walking on clouds" and store it in image_data.

Step 5: Save the generated image

Once we have the generated image data, we can save it to a file:

with open("test.png", "wb") as f:

This will save the generated image as test.png in the current directory.


In this example, we demonstrated how to use fal-serverless to generate images using the stable diffusion model. By utilizing fal-serverless infrastructure, we can easily deploy and scale this image generation process without worrying about managing servers or other infrastructure components.