Effortlessly Deploy Stable Diffusion on Google Cloud

Unleash the power of AI image generation with Stable Diffusion on Google Cloud. This guide provides a step-by-step tutorial on deploying Stable Diffusion using virtual machines on Google Cloud, enabling efficient and cost-effective AI art creation. Discover the advantages of cloud deployment over local setups and optimize your GPU resources for seamless image generation.

Why Deploy Stable Diffusion on the Cloud?

Deploying Stable Diffusion on Google Cloud offers significant advantages over local deployment. Local setups often require a dedicated, high-end GPU, such as the Nvidia RTX 4090

. While powerful for gaming, these cards aren't optimized for AI inference, leading to slower performance compared to professional AI GPUs like the T4 or L4. Google Cloud provides access to these specialized GPUs, allowing for faster and more efficient AI computations.

Another key benefit is cost savings. Instead of investing in expensive hardware that might sit idle, you can leverage Google Cloud's on-demand compute resources. Simply spin up a virtual machine with the necessary GPU power when you need it and shut it down when you're done. This pay-as-you-go model eliminates the upfront costs and ongoing maintenance associated with local hardware.

Stable Diffusion is a potent tool for creating images, and cloud deployment makes it more accessible and cost-effective. By utilizing the cloud, users are able to take advantage of scalable computing, allowing them to produce AI-generated art with increased speed. By utilizing the cloud, even novice AI enthusiasts can access potent computing tools without substantial financial investment.

Let's further consider two main methods of leveraging Google Cloud for Stable Diffusion

Google Colab vs. Virtual Machines: Choosing the Right Approach

Google offers two primary methods for accessing GPU power: Google Colab and Compute Engine virtual machines.

Google Colab provides a free, easy-to-use environment for running Python code, including AI models. However, it has limitations. While it offers a 'one-click' deployment

, it requires waiting around 10 mins before each execution.

Compute Engine virtual machines, on the other hand, offer more control and dedicated resources. Once configured, they provide an always-on environment, eliminating the need to re-run setup scripts every time. They are more streamlined than Google Colab. This guide will focus on the virtual machine approach, as it offers a superior user experience for deploying Stable Diffusion.

The choice between Google Colab and virtual machines depends on your needs and technical expertise. For beginners or those with limited resources, Google Colab provides a valuable starting point. However, for serious AI art creators seeking consistent performance and greater control, virtual machines are the preferred option. Here is a comparison table of the two options:

Creating a Virtual Machine Instance

Let's walk through the process of creating and configuring a virtual machine on Google Cloud for Stable Diffusion.

Step 1: Navigate to Compute Engine

From the Google Cloud Console, navigate to Compute Engine > VM instances. Click Create Instance to begin the process.

Step 2: Configure the Instance

• Give your instance a descriptive name.
• Select a region close to your location to minimize latency. Availability for GPUs may vary based on location.
• Choose a machine configuration that includes a GPU. Click the GPU tab. Select L4 as the GPU type and 1 as the Number of GPUs.
• Increase the vCPU and Memory to 8 vCPU and 32 GB memory for more efficent processing.
• Other configuration options are up to the user's preferences.

Step 3: Configure the Boot Disk

• In the Boot disk section, click Change to select a different image.
• Choose Deep Learning on Linux image. This image comes pre-configured with many of the necessary dependencies for AI workloads.
• Increase the disk size to at least 200 GB to accommodate the Stable Diffusion model and datasets.

Step 4: Set up Firewall Rules

• Under the Firewall section, enable both Allow HTTP traffic and Allow HTTPS traffic. This will allow you to access the Stable Diffusion web interface from your browser.

Step 5: Create a Static External IP Address

• Open the Networking tab and click on Network Interfaces to click on Edit.
• Then on the bottom on the page, click on Add IP address. Input a name for the static IP, and select Static for the type.
• Click Done and then click Create to launch your instance.

With the virtual machine created, the next steps involve setting up the firewall to allow external access and connecting to the instance via SSH.

Connecting to the Virtual Machine via SSH

Once the instance is created, connect to it via SSH. This can be done directly from the Google Cloud Console.

• Navigate to Compute Engine > VM instances.
• Find your newly created instance and click the SSH button in the Connect column.

This will open a terminal window in your browser, allowing you to interact with the virtual machine's command line.

Authorize it when a pop-up appears.

Installing NVIDIA Drivers and CUDA

The Deep Learning on Linux image comes with most necessary drivers. However, you may need to install the NVIDIA driver.

To verify the drivers and cuda are installed, you can run the nvidia-smi to confirm.

If it prompts to install the NVIDIA driver, select 'yes' and wait for installation to finish. This process will take a few mins, then restart the VM.

If there are problems, copy and paste the below to install the NVIDIA Driver manually:

sudo apt update
curl -s -L "https://raw.githubusercontent.com/GoogleCloudPlatform/compute-gpu-installation/main/linux/install_gpu_driver.py" | sudo python3
sudo python3 install_gpu_driver.py

Install and Update cuDNN

The steps for cuDNN may not be already pre-installed with your desired version. Copy the lines below and follow the steps:

sudo apt-get update
sudo apt-get install -y install cudnn-local-repo-debian12-12.4.1.1_1.0-1_amd64.deb
sudo cp /var/cudnn-local-repo-*/cudnn-*keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install -y cuda

Then run the command: cat /usr/include/cudnn_version.h to confirm it's running

This should return an error and not a version, telling the user they do not have cuDNN installed correctly.

To fix this, enter the following copied commands:

wget https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-local-repo-debian12-12-4-1_525.105.17-1_amd64.deb
sudo dpkg -i cuda-local-repo-debian12-12-4-1_525.105.17-1_amd64.deb
sudo cp /var/cuda-local-repo-debian12-12-4-1/cuda-*keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install -y cuda

sudo apt-get install -y cudnn

The version you install is reliant on the number after debian- the current number is 12 in this example, so your version will change based on the image.

Lastly, to ensure it's correctly installed use the command cat /usr/include/cudnn_version.h and you should now see the version numbers!

Install and Activate Conda

Copy and paste the code below in the correct order. The first line downloads Conda, and the following lines use the tool:

wget https://repo.anaconda.com/archive/Anaconda3-2024.02-1-Linux-x86_64.sh
bash Anaconda3-2024.02-1-Linux-x86_64.sh
source ~/.bashrc
conda update --base --name defaults conda
conda create --name sd python=3.10
source activate sd

Type 'yes' and hit enter when prompted and hit enter on any other prompts it gives you.

To test if it is installed correctly, use conda --version. If it returns with numbers you will know it's worked.

Installing and Configuring Stable Diffusion

To install Stable Diffusion, clone the Stable Diffusion repository with:

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
cd stable-diffusion-webui

To enable access, you will need to update webui-user.sh and the parameters. To do this:

nano webui-user.sh

Add the code below, after the git pull section. This provides extra memory and enables external access.

export COMMANDLINE_ARGS="--xformers --listen --enable-insecure-extension-access --no-half-vae"

Press CTRL+S to save, and then press CTRL+X to exit the editor. Lastly, paste this code to get it up and running:

mkohup webui.sh

WebUI should be fully running when it shows an output that has the two points:

• Running on local URL: http://0.0.0.0:7860
• Running on public URL: http://xxxxxxxx.grad.live

Access the Stable Diffusion Web Interface

Copy and paste the External IP from the VM, along with the port :7860 For example: 34.135.169.157:7860

Use that link to access it on another tab in your browser!

Change to Chinese Language

Follow these steps to change the output to Chinese Language:

Step 1: Go to Civitai.com and select Models

Step 2: Pick your desired checkpoint file, and copy the ID after /models

Step 3: Enter this ID to download your models using this command:

wget https://civitai.com/api/download/models/XXXXX

Replace XXXXX with your respective number to successfully download the checkpoint file.

Step 4: Then, go to https://github.com/VinsonLaro/stable-diffusion-webui-chinese

and go to Extensions to click Install from URL. Here, input the GitHub link and click Install

Step 5: Return to Installed and click Apply and Restart UI. Afterwards, when it has loaded, click on Settings and change the dropdown that mentions Localization to 中文英语 . Make sure to click to Apply Settings followed by hitting the Reload UI button that pops up.

Factors Influencing the Cost

Running Stable Diffusion on Google Cloud involves several cost factors. GPU usage is the primary expense, as more powerful GPUs and longer running times increase costs. Storage costs for models and generated images also contribute, along with networking charges for data transfer. Choosing the right instance type and optimizing usage patterns is crucial for cost-effectiveness. Preemptible instances offer discounted rates but may be terminated with 24 hours notice.

• GPU Usage: The type and duration of GPU usage significantly impact costs. Different GPUs have varying hourly rates, and running Stable Diffusion for extended periods will increase expenses.
• Storage: Storing models, generated images, and other data incurs storage costs. Opting for cost-effective storage options like Cloud Storage and managing data efficiently can help reduce expenses.
• Networking: Data transfer between Google Cloud services and external networks can incur networking charges. Minimizing data transfer and optimizing network configurations can contribute to cost savings.