February 2023 Summaries

Artists are increasingly exploring AI as a new frontier for creativity, with Disco Diffusion being a popular tool to generate AI art through machine learning by using text prompts to create images. This process, which can range from a few minutes to an hour, requires significant GPU computing power, and platforms like RunPod offer a way to rent this capacity. To use Disco Diffusion on RunPod, users need to set up an account, load credits, select an instance type, and deploy a pod, with options for cheaper but interruptible spot instances or stable on-demand instances. Once the pod is initialized, users can connect to it and configure settings to start rendering their art, which supports queued prompts and multiple GPUs, allowing for flexibility and efficiency in processing. The platform's interface facilitates monitoring progress and troubleshooting through logs, providing an accessible way for artists to experiment with AI-generated art while offering room for future improvements based on user feedback.

The blog post provides a comprehensive guide on using the Fast Stable template for stable diffusion on Runpod, emphasizing that while the template is packaged by Runpod for ease of use, the team does not maintain its code. Users are advised that the template is incompatible with encrypted volumes and should ensure that the GPU/CPU utilization is at 0% before connecting to avoid errors. The template includes support for the AUTOMATIC1111 web UI and custom training models with Dreambooth, with instructions on connecting to Jupyter Lab to run the necessary notebooks. It also covers the process of selecting and downloading stable diffusion models, including the option to use ControlNet, and stresses the importance of setting a username and password for the web UI to enhance security. For additional assistance, users are encouraged to seek help from the Runpod Discord community or other stable diffusion communities.

TheLastBen has recently enhanced their fast stable diffusion template by incorporating the offset noise functionality into Dreambooth, which can be activated by starting the template, connecting to the Jupyter link, and setting the Offset_Noise variable to "True" while increasing the total steps by 10-20%. This adjustment is said to produce impressive outputs, particularly in generating very dark and very light images, which are typically challenging for stable diffusion. For those interested in the underlying research and visual demonstrations, further details are available through a research blog post and a YouTube video.

Runpod's Serverless platform enables the creation of scalable API endpoints, and the tutorial provided offers a step-by-step guide to developing a basic worker that determines if a number is even, subsequently turning it into an API endpoint. Users need a local environment with Python, Docker knowledge, and the ability to build Docker containers, with all code available in the IsEven repository. The tutorial outlines creating a function that processes job inputs and returns results, and introduces runpod-python as a wrapper to manage API inputs. After testing the API locally and ensuring its functionality, the tutorial advises packaging the function into a Docker container and pushing it to a repository like Docker Hub. The final steps involve using Runpod's serverless templates to create an API endpoint and deploying it, with the option to test it using cURL or other online tools.

Creating a serverless AUTOMATIC1111 endpoint with a custom model that can scale efficiently is achievable by following a detailed guide that outlines necessary pre-requisites such as having Docker and Git installed on your computer, along with sufficient disk space and internet speed. The process involves using a GitHub repository as a base, cloning it, and making minimal code changes to incorporate your custom model, either by adding it locally or downloading it via a public link. Once the Docker image is built and the start script is configured to point to the new model, the image can be pushed to a container registry like Docker Hub for public use. While deploying the API comes with conveniences, like extensive built-in functionalities, users should be aware of potential cold start times and ensure all features work as intended, making it an effective starting point without needing to develop a custom code base from scratch.