1. Workflow Overview

This workflow leverages Alibaba's Wanx-8G model for text-to-video generation, featuring:

• Beginner-friendly: Pre-configured parameters

• Advanced control: Supports LoRA fine-tuning & tiled decoding

• Multi-format output: Direct MP4 (H.264) or animated image export

2. Core Models

3. Key Nodes

4. Workflow Structure

• Group 1: Text Input

  • CLIPTextEncode: Processes positive (e.g., "A fox in snowy scenery") and negative prompts

• Group 2: Model Loading

  • UNETLoader: Loads Wanx-8G main model

  • LoraLoaderModelOnly: Optional LoRA (default strength=0.8)

• Group 3: Video Generation

  • KSampler: Uses UniPC sampler (30 steps, CFG=6)

  • VAEDecodeTiled: Decodes latent with tiling

5. Inputs & Outputs

• Required Inputs:

  • Positive prompt (English/Chinese)

  • Negative prompt (pre-set quality filters)

  • Frame count (default=33)

• Outputs:

  • MP4 video (16FPS, H.264)

  • Resolution: 832x480 (adjustable)

6. Notes

⚠️ Critical Configs:

1. VRAM Optimization:

   • Enable VAEDecodeTiled for 8GB GPUs

   • Use uni_pc sampler for faster generation

2. Plugin:

   git clone https://github.com/AI-ModelScope/comfyui-hunyuan-plugin

3. Video Quality:

   • Adjust crf in VHS_VideoCombine (18-28, lower=better)