MeiGen-MultiTalk

MeiGen-MultiTalk Endpoint Handler

This repository contains a custom handler for deploying MeiGen-AI's MultiTalk model on Hugging Face Inference Endpoints.

Model Description

MeiGen-MultiTalk is an advanced model for generating audio-driven multi-person conversational videos. This handler wraps the original model to work with HF Inference Endpoints.

Features

• Text-to-video generation
• Image-to-video generation
• Multi-person conversation synthesis
• Support for various resolutions (480p, 720p)
• Optimized for A100 GPUs

Usage with Inference Endpoints

Recommended Configuration

• Hardware: GPU · A100 · 1x GPU (80 GB)
• Autoscaling:
  • Min replicas: 0
  • Max replicas: 1
  • Scale to zero after: 300 seconds

API Example

import requests
import json
import base64

API_URL = "https://YOUR-ENDPOINT-URL.endpoints.huggingface.cloud"
headers = {
    "Authorization": "Bearer YOUR_HF_TOKEN",
    "Content-Type": "application/json"
}

# Text-to-video generation
data = {
    "inputs": {
        "prompt": "A person giving a presentation"
    },
    "parameters": {
        "num_frames": 16,
        "height": 480,
        "width": 640,
        "num_inference_steps": 25,
        "guidance_scale": 7.5
    }
}

response = requests.post(API_URL, headers=headers, json=data)
result = response.json()

Technical Details

The handler includes:

• Automatic model loading from MeiGen-AI/MeiGen-MultiTalk
• Memory optimization for GPU inference
• Support for both diffusion pipeline and transformer modes
• Error handling and logging
• Base64 encoding for image/video I/O

License

Apache 2.0 License

Credits

Based on the original MeiGen-AI/MeiGen-MultiTalk model.