GenAI 3D Asset Generation Demo

This project demonstrates a pipeline for generating 3D assets using generative AI. It consists of three main components: a frontend application, a worker pool for processing, and a 3D model generation service.

Architecture Overview

The architecture is composed of three main services:

1. Frontend Application: A web interface for users to submit requests for 3D asset generation.
2. Worker Pool: A scalable pool of workers that queue and manage asset generation tasks.
3. Hunyuan3D Service: A GPU-accelerated service that generates the 3D models.

These components communicate with each other and utilize shared storage for assets and models.

Components

1. Hunyuan3D Service

This service is responsible for the core 3D model generation.

• Source Directory: hunyuan3D/
• Dockerfile: hunyuan3D/Dockerfile (based on an NVIDIA image)
• Deployment: Cloud Run with GPUs

2. Frontend Application

The user-facing web application.

• Dockerfile: Dockerfile-frontend
• Deployment: Cloud Run

3. Worker Pool

This service manages the queue of generation tasks submitted by the frontend.

• Dockerfile: Dockerfile-worker
• Deployment: Cloud Run

Setup and Deployment

MUST HAVE

• NFS mounted storage should point to /root and already contain all the models from hugging face, in particular:

    -- 7.9G    .cache/huggingface/hub/models--tencent--Hunyuan3D-2mini
    -- 18G     .cache/huggingface/hub/models--tencent--Hunyuan3D-2

Prerequisites

• Google Cloud Project with billing enabled.
• gcloud CLI installed and authenticated.
• Docker installed. (for local builds, so it's optional)
• A shared NFS folder (e.g., Filestore) for the Hunyuan3D service as well as shared models and image assets

Environment Variables

It is crucial to configure the environment variables correctly. The most important one is passing the URL of the deployed Hunyuan3D service to the Worker Pool application.

• HUNYUAN3D_URL: The endpoint of the Hunyuan3D service. This needs to be set for the worker service. (don't forget the /generate path)
• VERTEX_REGION: In which region you want to invoke vertex services such as imagen
• PROJECT_ID: Your Google Cloud project ID.
• TOPIC_NAME: The name of the pubsub topic to publish to.
• SUBSCRIPTION_NAME: The name of the pubsub subscription to listen to.

Deployment Steps

1. Set up Shared Storage

1. Create a Filestore instance for the NFS share required by the Hunyuan3D service. Note the NFS share IP address and file share name.
2. Create a GCS bucket that will be used by the frontend and worker services. It needs to be mounted at /app/static/models.

2. Deploy Hunyuan3D Service

You can deploy this service to either Cloud Run with a GPU or GKE Autopilot with a GPU.

** Deploy to Cloud Run**

Create a service with GPU support and 8vCPU / 32GB of RAM. Mount Filestore so it becomes /root in the container.

After deployment, get the external IP/domain of the service and set it as HUNYUAN3D_URL.

3. Deploy Frontend and Worker

Frontend (Docker-frontend) is deployed as Cloud Run service and Worker (Docker-worker) is deployed as a Worker Pool in Cloud Run. Make sure they have the right permissions for Pub Sub, Cloud Storage and Vertex AI.