README.md

---
title: Avatar Demo
emoji: 🎬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
python_version: "3.10"
app_file: app.py
hardware: zero-a10g
suggested_hardware: zero-a10g
startup_duration_timeout: 1h
license: apache-2.0
short_description: Text, image, and voice to talking avatar video
models:
  - nineninesix/gepard-1.0
  - nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps
  - microsoft/wavlm-base-plus
  - meituan-longcat/LongCat-Video
  - meituan-longcat/LongCat-Video-Avatar-1.5
tags:
  - text-to-speech
  - talking-avatar
  - video-generation
  - zerogpu
---

# Avatar Demo

This Space merges GEPARD TTS with LongCat-Video-Avatar 1.5 in one Gradio app.

Workflow:

1. Enter text.
2. Upload a reference face image.
3. Pick a GEPARD preset voice or upload a voice reference clip.
4. Generate speech audio and a 5-second lip-synced avatar video.

The GEPARD checkpoint is private, so the Space must have an `HF_TOKEN` secret
with access to `nineninesix/gepard-1.0`.

ZeroGPU notes:

- `create_env.py` re-pins `transformers==5.3.0` at startup before ML imports.
- `spaces` is imported before `torch`.
- Both model stacks are loaded at module level and moved to `cuda`.
- The generation handler is wrapped with `@spaces.GPU(..., size="xlarge")`
  using a duration callable that returns 300 seconds.
- LongCat's xformers call sites use a local PyTorch SDPA shim.