Wat3r B3nding (Awan Effect) LoRA for WAN 2.2
This LoRA is fine-tuned for the WAN 2.2 video model to generate a specific, dynamic liquid transformation effect, transitioning a subject from one state or appearance to an entirely new one via a massive surge of hyper-realistic water.
🌊 Model Summary
- Goal: To implement a stylized, fluid transition (the "Awan Effect") between two different subjects or scene compositions.
- Base Model: Designed to be used with the WAN 2.2 video models (e.g., WAN2.2-T2V-A14B or I2V variants).
- Unique Token (Trigger Word):
wat3r b3nding
🚀 Usage
Recommended Prompt Structure
The LoRA works best with highly descriptive captions that specify both the starting subject/scene and the ending subject/scene, framing the transition with the trigger word.
| Component | Example Detail |
|---|
| Starting Scene | A subject with long red curly hair and visible arm and chest tattoos, seated and smiling in a retro diner booth. |
| Transition Token | wat3r b3nding |
| Ending Scene | rising into a fluid human silhouette of a different subject wearing a high-collared, form-fitting blue sequined qipao-style dress with cutouts, surrounded by a swirling flock of black and white magpies. |
| Style/Descriptors | Dynamic, flowing, and elemental. |
Example Prompt
A subject with long red curly hair and visible arm and chest tattoos, seated and smiling in a retro diner booth, Wat3r B3nding, then rising into a fluid human silhouette of a different subject wearing a high-collared, form-fitting blue sequined qipao-style dress with cutouts, surrounded by a swirling flock of black and white magpies. Dynamic, flowing, and elemental.
🎬 Video Examples
Example 1
Example 2
Example 3
Example 4
Example 5
Example 6
Example 7
Example 8
Example 9
Example 10
Example 11
Example 12
🛠️ Training Details
These settings were optimized for a small training dataset (3 videos/subjects) to prevent overfitting to the subjects and focus solely on the water-bending effect.
| Setting | Value/Strategy | Notes |
|---|
| Trigger Word | wat3r b3nding | Use a unique token to minimize interference from the base model's knowledge. |
| Dataset Size | 3 Video Clips | |
| Target Steps | 1100 Total Steps | |
| Network Rank (Dimension) | 32 | |
| Alpha | 32 | |
| Learning Rate (LR) | $1e-4$ | A very low LR is crucial for small datasets to ensure stable training. |
| Optimizer | AdamW | AdamW is the standard, reliable choice. |
| Precision | BF16 | Required for memory efficiency on larger models. |
