1. Workflow Overview

• Purpose: Generates videos from text (T2V) or images (I2V) using Tencent's WanXiang 1.3B model, with post-processing like super-resolution and frame interpolation.

• Core Models:

  • Wan2_1-T2V-1.3B: Text-to-video base model (1.3B params).

  • Wan2_1-I2V-14B: Image-to-video model (14B params, FP8 quantized).

  • UMT5-XXL: Multilingual text encoder (supports Chinese prompts).

  • FILM-Net: Frame interpolation model (for smoother videos).

2. Key Nodes

3. Workflow Structure

• Group 1: Text-to-Video (T2V)

  • Input: Prompts (e.g., "vibrant cartoon style"), negative prompts, seed.

  • Process: Text encoding via UMT5-XXL, video generation by Wan2_1-T2V.

  • Output: Raw video (720P).

• Group 2: Image-to-Video (I2V)

  • Input: Reference image (e.g., ComfyUI_06397_.png), prompts.

  • Process: Uses CLIP-Vision for image features, Wan2_1-I2V for generation.

  • Output: Dynamic video (81+ frames).

• Group 3: Post-Processing

  • Upscaling: ESRGAN_4x for higher resolution.

  • Frame Interpolation: Boosts FPS from 16 to 32 via FILM-Net.

  • Output: Final MP4 (H.264, CRF=19).

4. Input & Output

• Key Inputs:

  • Prompts: Mix Chinese/English (e.g., "产品摄影，拉近镜头").

  • Resolution: Default 720x1280 (portrait) or 1280x720 (landscape).

  • Frames: I2V requires ≥81 frames.

• Output:

  • Format: MP4 (H.264).

  • Path: Saved to ComfyUI/output/Hunyuan/videos/.

5. Notes

• VRAM Requirements:

  • I2V model needs ≥16GB GPU (FP8 quantization still demands high-end cards).

  • Enable torch.compile for speed (requires CUDA 12+).

• Common Issues:

  • Frame count <81: Adjust WanVideoBlockSwap settings.

  • OOM errors: Reduce resolution or disable bf16.

• Optimization:

  • Use TeaCache node to save VRAM (enabled by default).

  • Include motion keywords in prompts (e.g., "zoom in").