# AI Avatar Video from Voice Using Just 10GB VRAM (WanGP 5.4 + Hunyuan)

**URL:** <https://onehack.st/t/ai-avatar-video-from-voice-using-just-10gb-vram-wangp-5-4-hunyuan/306342>\
**Category:** Tools & Scripts\
**Tags:** tools, ai\
**Created:** [June 10, 2025, 2:32am UTC](https://onehack.st/t/ai-avatar-video-from-voice-using-just-10gb-vram-wangp-5-4-hunyuan/306342 "2025-06-10T02:32:36Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![SRZ](https://onehack.st/user_avatar/onehack.st/srz/32/146132_2.png) [@SRZ](https://onehack.st/u/SRZ)\
**Post date:** [June 10, 2025, 2:32am UTC](https://onehack.st/t/ai-avatar-video-from-voice-using-just-10gb-vram-wangp-5-4-hunyuan/306342/1 "2025-06-10T02:32:36Z")

</div>

### Short Introduction

You can now create highly realistic, lip-synced talking avatar videos — powered by AI — using nothing more than a standard 10GB VRAM GPU. With WanGP 5.4 and the Hunyuan Video Avatar model, anyone can generate stunning 15-second clips that match speech or songs to facial expressions and motion, all locally and without needing enterprise hardware.

* * *

### Simplified One-Line Flowchart

Voice/song + Portrait Image ➔ WanGP Web UI + Hunyuan Model ➔ 15s Talking Avatar Output

> <https://x.com/deepbeepmeep/status/1930672881320460357>

* * *

### Easy Step-by-Step Method

**Step 1: Get the Tool**

- Install from GitHub:  
[https://github.com/deepbeepmeep/Wan2GP](https://github.com/deepbeepmeep/Wan2GP)
- This tool is browser-based, lightweight, and supports multiple video models including Hunyuan.

**Step 2: Prepare Inputs**

- **Voice or song clip** up to 15 seconds for best stability.
- **Portrait image** resized to exactly **480x832** to prevent tensor mismatch errors during generation.

**Step 3: Launch the Interface and Set Parameters**

- Open WanGP in your browser.
- Upload your voice/audio file and image.
- Select **Hunyuan Video Avatar** as your model.
- Recommended settings:
  - Steps: 10–30 (20 is a good balance)
  - Resolution: 512x512
  - Attention: `auto` or `sage2`
  - Data Type: `BF16`
  - Quantization: `Scaled`

**Step 4: Optimize for Smooth Generation**

- If generation is sluggish or fails:
  - Use **PyTorch 2.7.1 with video/audio support**
  - Add `triton` and `sage-attention`
  - Avoid uploading full-length songs; trim to small chunks (5–10s)

**Step 5: Optional Add-ons & Advanced Use**

- GGUF-prepped model version for manual handling:  
[https://huggingface.co/lym00/HunyuanVideo-Avatar-GGUF](https://huggingface.co/lym00/HunyuanVideo-Avatar-GGUF)
- If installing via Docker, you can change ports or patch DNS issues manually:

---

<div class="post-metadata">

**Author:** ![TheJoker](https://onehack.st/user_avatar/onehack.st/thejoker/32/22660_2.png) [@TheJoker](https://onehack.st/u/TheJoker)\
**Post date:** [June 10, 2025, 2:33am UTC](https://onehack.st/t/ai-avatar-video-from-voice-using-just-10gb-vram-wangp-5-4-hunyuan/306342/2 "2025-06-10T02:33:32Z")

</div>