wangkanai / wan-lightx2v

WAN LightX2V LoRA Adapters

This repository contains LoRA (Low-Rank Adaptation) adapters for the LightX2V Text-to-Video model. These adapters enable efficient fine-tuning and inference with reduced computational requirements.

📦 Model Information

• Base Model : LightX2V T2V 14B
• Type : CFG Step Distillation LoRA Adapters
• Version : v2
• Precision : BF16 (Brain Floating Point 16)
• Available Ranks : 8, 16, 32, 64, 128

📁 Repository Structure

loras/
└── wan/
    ├── wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank8-bf16.safetensors
    ├── wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank16-bf16.safetensors
    ├── wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank32-bf16.safetensors
    ├── wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank64-bf16.safetensors
    └── wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank128-bf16.safetensors

🎯 LoRA Rank Selection Guide

Choose the appropriate rank based on your requirements:

🚀 Usage

Using with Diffusers

from diffusers import DiffusionPipeline
import torch

# Load base model
pipe = DiffusionPipeline.from_pretrained(
    "lightx2v/lightx2v-t2v-14b",
    torch_dtype=torch.bfloat16
)

# Load LoRA adapter (example with rank 32)
pipe.load_lora_weights(
    "your-username/wan-lightx2v",
    weight_name="loras/wan/wan-lightx2v-t2v-14b-cfg-step-distill-v2-rank32-bf16.safetensors"
)

# Generate video
prompt = "A beautiful sunset over the ocean with waves crashing"
video = pipe(prompt).frames

# Save video
# ... your video saving code here

Using with ComfyUI

1. Download the desired LoRA file
2. Place it in your ComfyUI/models/loras/ directory
3. In your workflow, add a "Load LoRA" node
4. Select the corresponding LoRA file
5. Adjust the strength (recommended: 0.8-1.0)

Using with Other Frameworks

These .safetensors files are compatible with any framework that supports LoRA loading for diffusion models.

⚙️ Technical Details

CFG Step Distillation

These LoRAs utilize Classifier-Free Guidance (CFG) step distillation, which:

• Reduces the number of inference steps required
• Maintains or improves generation quality
• Accelerates video generation process
• Optimizes the guidance scale behavior

BF16 Precision

The adapters use Brain Floating Point 16 (BF16) format:

• Better numerical stability than FP16
• Wider dynamic range
• Recommended for modern GPUs (Ampere architecture and newer)
• Compatible with mixed-precision training and inference

📊 Performance Benchmarks

(Add your benchmark results here if available)

🔧 Requirements

• Python 3.8+
• PyTorch 2.0+ (with BF16 support)
• Diffusers library
• GPU with BF16 support (NVIDIA Ampere or newer recommended)
• Sufficient VRAM based on chosen rank

📝 Citation

If you use these LoRA adapters in your research or projects, please cite:

@misc{wan-lightx2v-lora,
  title={WAN LightX2V LoRA Adapters},
  author={[Your Name]},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/[your-username]/wan-lightx2v}}
}

📄 License

(Specify your license here - e.g., MIT, Apache 2.0, Creative Commons, etc.)

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page.

📧 Contact

(Add your contact information or links to your social media/website)

🙏 Acknowledgments

• LightX2V team for the base model
• Community contributors and testers

---

Note : This is a model repository. Please ensure you have the necessary computational resources and comply with the base model's license terms when using these adapters.