Stylish Logo Creation with Stable Diffusion on Vultr Cloud GPU

Updated on August 28, 2023
Stylish Logo Creation with Stable Diffusion on Vultr Cloud GPU header image


The emergence of image-generative solutions has revolutionized the way of perceiving and interacting with visual content. It harnesses the power of artificial intelligence to create stunning, lifelike images that span a spectrum of artistic styles and creative visions. Stable Diffusion, as one of the main players in this field, offers open-sourced solutions for new forms of image expression and design.

When using Stable Diffusion on a GPU-powered server, it's possible to create stunning logo or banner designs without any graphic design skills. In this article, create stylish logos on a Vultr Cloud GPU instance, export, and share them for offline usage on your local computer.


Before you begin:

  • Deploy a Debian Vultr Cloud GPU Serverwith at least:

    • 1/7 GPU
    • 10 GB GPU RAM
    • 2 vCPUs

    If you deploy an instance with less than 10 GB of GPU RAM, you may encounter errors like RuntimeError: CUDA error: out of memory when loading the models in this article.

  • Using SSH, access the server

  • Create a non-root user with sudo privileges

  • Switch to the user

      # su user_example

Install Python Libraries

To generate visual content, install necessary Python libraries required by the Stable Diffusion model, and a development environment such as Jupyter Notebook to execute Python code. Install all required model dependencies as described in the steps below.

  1. Update the Pip package manager

     $ pip install --upgrade pip
  2. Using pip, install the PyTorch package

     $ pip install torch
  3. Install diffusers, opencv, accelerate, transformers, and xformers packages

     $ pip install diffusers opencv-python accelerate transformers xformers

    Below is the function of each package:

    • opencv: Allows you to perform image processing and computer vision tasks. In this article, canny inside opencv detects logo edges
    • accelerate: Provides faster and less memory-intense model loading
    • transformers,xformers: Enable the creation of a StableDiffusionControlNetPipeline
  4. Install Jupyter Notebook

     $  pip install notebook
  5. Allow the Jupyter Notebook port 8888 through the default UFW firewall

     $ sudo ufw allow 8888/tcp
  6. Start Jupyter Notebook

     $ jupyter notebook --ip=

    The above command starts Jupyter Notebook listening for incoming connections on all server interfaces as declared with If the above command fails to run, stop your SSH session, and re-establish a connection to the server.

    When successful, an access token displays in your output as below:

     [I 2023-08-16 20:42:14.446 ServerApp] http://HOSTNAME:8888/tree?token=c70592882983e614515ee7ed3fbbcfc87b6a8d88d4a79a4e
     [I 2023-08-16 20:42:14.446 ServerApp]
     [I 2023-08-16 20:42:14.446 ServerApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
     [W 2023-08-16 20:42:14.449 ServerApp] No web browser found: Error('could not locate runnable browser').
     [C 2023-08-16 20:42:14.449 ServerApp] 
         To access the server, open this file in a browser:
         Or copy and paste one of these URLs:

Start the Logo Generation Process

  1. Using a web browser such as Chrome, access your Jupyter Notebook interface using the generated access token
  2. Within the Jupyter interface, click New, and select Notebook from the dropdown list to create a new Python3 Kernel file

  3. In the new Notebook code cell, update Jupyter and Ipywidgets

     pip install --upgrade jupyter ipywidgets
  4. Press Ctrl + Enter to run the command and other code in the file

  5. Import required libraries as below

     import torch
     import cv2
     import numpy as np
     from PIL import Image
     from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
     from diffusers.utils import load_image
  6. To load a base image to the model, define a new img_url variable with your target image URL, and use the Stable Diffusion load_image utility to load the image as below

     img_url = ""
     image = load_image(img_url)

    In the above code, the variable uses a Vultr logo. Replace the link with the actual original logo URL you intend to style.

Apply the OpenCV Canny Algorithm

  1. Load the image into an OpenCV-compatible input in a NumPy array

     image = np.array(image)
  2. Apply canny to detect and highlight the edges of the base logo image

     low_threshold, high_threshold = 100, 200
     image = cv2.Canny(image, low_threshold, high_threshold)[:, :, None]
     image = np.concatenate([image, image, image], axis=2)
     image = Image.fromarray(image)
  3. Save and view canny processing output"canny.png")

    When successful, your base logo image should display with highlighted edges

    Output after applying canny

Create a Stable Diffusion Pipeline using ControlNet

  1. Declare the base Stable Diffusion model path

     sd_model_path = "danbrown/RevAnimated-v1-2-2"

    The example in this article uses the HuggingFace Rev Animated model. Different models generate images in different styles, it's recommended to explore other models to generate different results.

  2. Declare the ControlNet model path

     controlnet_model_path = "lllyasviel/sd-controlnet-canny"
  3. Load the model in half precision

     controlnet = ControlNetModel.from_pretrained(controlnet_model_path, torch_dtype=torch.float16)
  4. Create the Stable Diffusion pipeline

     pipe = StableDiffusionControlNetPipeline.from_pretrained(
     pipe ="cuda")

Setup the Pipeline parameters

  1. Enable xFormers memory efficient attention to lower GPU memory usage and speed up the inference process

  2. Offload all models to the CPU to reduce GPU memory usage with a low impact on the image generation process

  3. Declare the pipeline scheduler

     pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)

    The above code defines the method to update a sample image based on model outputs. The UniPC scheduler is fast and enables a reasonable generation performance.

  1. Define the prompt and the negative prompt. The more detailed the prompt is, the better the generated image becomes

     prompt = "best quality, masterpiece, sharp silver constructions, spaceship, Mars, crater, universe background"
     negative_prompt = "lowres, cropped, worst quality, low quality, watermark, blurry, EasyNegative"
  2. Generate the stylish logo image with proper parameters. num_inference_steps, guidance_scale, and controlnet_conditioning_scale are the most important among many others.

     output = pipe(

    Below is what the above code parameters do:

    • num_inference_steps: Defines the number of inference steps during the diffusion process. The value cannot be too small as the diffusion process requires enough steps to generate the image. As a result, 30 steps balance the image quality and the computational resource used.
    • guidance_scale: Controls how strict the diffusion process needs to follow the given prompts. The value 7.0 offers a good balance between creative freedom and your defined direction.
    • controlnet_conditioning_scale: Determines how obvious the control image appears in the generated image. A small value makes the control image more invisible and merged into the whole image. 0.8 is an experimental value, it's recommended to try different values to test the generation differences.

    The generated styled logo image should look look like if not fully match your text prompt

    Space themed Image

Save the Generated Image

  1. Save the generated Stylish Logo Image to an existing server directory"/home/user/image.png")

    The above code exports the generated image to your user's home directory. To define a custom directory, verify that it exists before running the code

  2. To view the saved images, navigate to your main Jupyter interface, double-click the file to open and view the exported image

    View Generated Images in Jupyter

    To download the generated images from your server to your local machine, use a secure file transfer protocol such as SFTP, SCP, or Rsync to save a copy of the files locally.

Refine the Generated Logo Images

When running the seed, the pipeline is not fixed, as such you can try different prompts and parameters to refine your generated stylish logo image. For example, change the prompt as below

prompt = "best quality, masterpiece, tropical islands spread in blue ocean, blue sky and white cloud, soft sunshine"

The above prompt generates a different image based on the same input image. You can apply more prompts to refine your output and style your logo image even further

Beach themed Image


In this article, you created a stylish logo using Stable Diffusion on a Vultr cloud GPU instance. Using Stable Diffusion allows you to generate stunning logo or banner images in minutes which improves work efficiency. For more information on how to use Stable Diffusion, visit the model FAQ page.

Next Steps

To implement more solutions on your Vultr Cloud GPU Server, visit the following resources: