1. Workflow Overview

• Purpose: Generates dynamic flower bloom effects on buildings using Tencent's WanXiang 2.1 model (I2V), with super-resolution and frame interpolation.

• Core Models:

  • Wan2_1-I2V-14B-480P_fp8: Image-to-video model (FP8 quantized, 480P output).

  • UMT5-XXL: Multilingual text encoder.

  • 4xFaceUpSharpDAT: Upscaler optimized for plant/architectural details.

  • FILM-Net: Frame interpolation model.

2. Key Nodes

3. Workflow Structure

• Group 1: Input Setup

  • Image Input: LoadImage node loads building image (e.g., output (13).png).

  • Text Prompts: Use StringConstantMultiline for descriptions (e.g., "Vines climbing buildings in a continuous frame").

  • LoRA: Activates Blooming LoRA for enhanced plant growth effects.

• Group 2: Video Generation

  • Model: Wan2_1-I2V-14B at 480P with FP8 quantization.

  • Frame Control: WanVideoBlockSwap sets frame count (81 frames recommended).

  • Sampler: dpm++ with CFG=5.

• Group 3: Post-Processing

  • Upscaling: 4xFaceUpSharpDAT for detail enhancement.

  • Interpolation: FILM-Net doubles frame rate to 32fps.

  • Output: MP4 (H.264, CRF=19).

4. Input & Output

• Key Inputs:

  • Image: Recommended size 480x768 (portrait) or 768x480 (landscape).

  • Prompts: Include motion cues (e.g., "flowers gradually covering walls").

  • LoRA Weight: Default 1.0 (adjustable in WanVideoLoraSelect).

• Output:

  • Format: MP4 (AnimateDiff_xxxx.mp4).

  • FPS: 16 (raw) or 32 (interpolated).

5. Notes

• VRAM Requirements:

  • ≥12GB for base generation, ≥16GB with upscaling.

  • FP8 reduces VRAM usage but may lose details.

• Troubleshooting:

  • Weak bloom effect: Increase LoRA weight or prompt details.

  • Choppy video: Check FILM VFI multiplier (keep ≤2x).

• Optimization:

  • Enable TeaCache in WanVideoSampler for VRAM efficiency.

  • Pre-process images to remove cluttered backgrounds (use LayerUtility).