The Livepeer AI Video Subnet is in its Alpha phase. Bugs or issues may be encountered. Contributions to improvement are appreciated - please report problems via the issue tracker. Feedback is invaluable for enhancing the AI Subnet. Thank you for your contributions!

The AI Video Subnet, also known as the AI Subnet, is the first step toward bringing powerful AI video capabilities into the Livepeer network. It enables video developers to add a rapidly growing suite of generative AI features such as text-to-image, image-to-image, image-to-video and upscaling to their applications. Livepeer Node operators are able to earn revenue by deploying their GPU resources for AI processing tasks. Ready to dive in? Choose one of the cards below to kickstart your journey with the AI Subnet.

Kickstart Your Journey

Background on the AI Subnet

The AI Subnet, initially proposed in this SPE treasury proposal, represents a significant evolution within the Livepeer ecosystem. This decentralized, open-source framework seamlessly integrates a variety of generative image and video AI inference tasks into the existing Livepeer network infrastructure. These enhancements strengthen Livepeer’s transcoding services, celebrated for their low cost and high reliability, while also paving the way for groundbreaking applications across both emerging Web3 environments and established Web2 sectors.

Designed to revolutionize creative processes, the AI Subnet reflects Livepeer’s commitment to creating a globally accessible open video infrastructure. Livepeer’s AI subnet aims to empower users to leverage any form of image and video computing with AI. By equipping video applications with advanced AI tools and reducing reliance on centralized computing resources, we are dedicated to extending cutting-edge AI capabilities to a broader audience, fostering a more equitable digital landscape.

Advantages of Livepeer’s AI Subnet

  • Decentralization: Enhances security and resilience by eliminating single points of failure.
  • Cost-Effectiveness: Expected to provide AI inference capabilities at a significantly lower cost compared to traditional cloud services.
  • Scalability: Easily scales in response to user demand, ensuring reliable service without interruptions.
  • Open-Source Innovation: Encourages collaborative development, accelerating innovation and broadening the range of supported applications.

How It Works

The AI Subnet, built on the established Livepeer network, leverages a decentralized payment infrastructure for compensating AI orchestrator nodes for performing AI inference tasks. The network consists of two primary actors:

  • AI Gateway Nodes: These nodes manage the flow of AI tasks from applications and users, directing them to the appropriate Orchestrator nodes based on capability and current load, ensuring efficient task allocation and system scalability.

  • AI Orchestrator Nodes: These nodes manage the execution of AI tasks. They can perform AI inference tasks on GPUs located on the same machine or be connected to multiple machines with multiple GPUs, known as AI Worker Nodes. The orchestrator or its worker nodes keep AI models warm on their GPUs for immediate processing and can dynamically load models as tasks arrive, optimizing both response time and resource utilization.

This architecture is designed for scalability, enabling easy integration of additional Orchestrator and Gateway nodes as demand increases. Under the hood, it relies on a specialized ai-runner Docker image to execute inference requests on AI models, simplifying deployment and enhancing the scalability of new pipelines. Ongoing developments aim to enhance performance and broaden the container’s capabilities to support increasingly complex AI models and custom user-defined pipelines.

Below is a simplified diagram illustrating the complete AI inference pipeline within the Livepeer AI Subnet. Only one AI Orchestrator with one GPU on the same machine is shown for clarity. In reality, the AI Gateway is connected to multiple Orchestrators, each of which can have multiple worker nodes with various GPUs attached.

This flow starts at the AI Gateway nodes, which route tasks from applications or users to an appropriate AI Orchestrator node. The selection of the AI Orchestrator node is based on factors such as the speed of previous inference requests, orchestrator stake, advertised price, and more. The AI Orchestrator node then executes the task in the ai-runner Docker Container. In this container, the AI Orchestrator can either:

  • Pre-load a model: Orchestrators keep frequently used models warm on GPUs, speeding up task processing.
  • Dynamically load a model: Orchestrators can load models on demand, allowing flexibility to handle various tasks.

During the selection process, to ensure quick response times, Gateways first check if there are any Orchestrators available with the requested model ‘warm’ on their GPU. If not, they route the task to an Orchestrator that can dynamically load the model. Once the model is loaded, the task is executed on the GPU, and the results are returned to the Orchestrator node, which then sends them back to the Gateway node and finally to the application or user that requested the task. This architecture ensures the AI Subnet is both efficient and flexible, ready to scale and adapt to various AI applications.

Explore Generative AI Pipelines

The AI Subnet hosts a variety of generative AI pipelines, each supporting different models. Current offerings primarily utilize Diffusion models. Plans are in place to expand support to include other model types in future updates.

AI Pipelines

Dive into the AI Pipelines page to learn more about each pipeline and the models they support.

Current Limitations and Future Directions

  • Alpha Phase: The AI Subnet is currently in its Alpha phase, and users may encounter bugs or issues during this early stage. It is not yet meant to be used with high demand production workloads.
  • Supports Limited Set of Open-source Models: The AI Subnet currently supports a limited set of AI models and pipelines that are open-source and available on Hugging Face. However, this range is gradually expanding with the goal of supporting any custom model in the future.
  • Only Higher VRAM GPUs Supported: Currently, the AI Subnet requires GPUs with at least 16GB of VRAM to run AI inference tasks effectively. We are working to expand this support to lower VRAM GPUs in the future.

AI Subnet Terminology

  • Mainnet Transcoding Network: Comprises Orchestrator and Gateway nodes that perform or coordinate transcoding tasks.
  • AI Subnet (also known as AI Video Subnet): A specialized subnet within the Livepeer network designed for managing AI inference tasks.
  • Mainnet Transcoding Network Orchestrator: A node that handles transcoding tasks within the Mainnet Transcoding Network. Often referred to as Orchestrator.
  • Mainnet Transcoding Network Gateway: A node that routes transcoding tasks to the appropriate Orchestrator nodes. Often referred to as Gateway and formerly known as Broadcaster.
  • AI Orchestrator: A specialized node responsible for executing AI inference tasks within the AI Subnet.
  • AI Gateway: A specialized node that directs AI tasks to the appropriate AI Orchestrator nodes within the AI Subnet.