What is ControlNet

ControlNet enhances image-to-image generation by conditioning input images, enabling versatile control using techniques like pose and edge detection.

Imagine having a striking image with great composition, but desiring to refill it with different content. This is where ControlNet excels. Let's dive into its capabilities with an example:

In this example, the image on the left side is the original image from the Internet. In the middle, it is the depth structure of the image, produced by Hyperbolic AI preprocessor. On the right, the image is generated using SDXL + ControlNet, seamlessly integrating depth information and the text prompt of "an astronaut on Mars". To explore this firsthand, utilize the following Python code:

import base64
from io import BytesIO

import requests
from diffusers.utils import load_image
from PIL import Image

def encode_image(img):
    "Encode PIL.Image into base64 string"
    buffered = BytesIO()
    img.save(buffered, format="PNG")
    encoded_string = base64.b64encode(buffered.getvalue()).decode("utf-8")
    return encoded_string

def decode_image(encoded_string):
    "Decode base64 string into PIL.Image"
    image_data = base64.b64decode(encoded_string)
    image = Image.open(BytesIO(image_data)).convert("RGB")
    return image

image = load_image(


headers = {
    'Authorization': f'Bearer {token}',

result = requests.post(
        "prompt": "an astronaut on Mars",
        "backend": "auto",
        "model_name": "SDXL-ControlNet",
        "height": 1024,
        "width": 1024,
        "controlnet_name": "depth",
        "controlnet_image": encode_image(image),
        "seed": 5742320,
        "cfg_scale": 15,

image_str = result.json()["images"][0]["image"]

img = decode_image(image_str)

Currently, HyperbolicAI supports two Stable Diffusion + ControlNet: SDXL-ControlNet and SD1.5-ControlNet. Furthermore, there are four distinct types of ControlNet available: canny, depth, openpose and softedge.