Deploy WAN LoRA Training

Fine-tune WAN (text-to-video) with LoRA using your own short clips and images.

🚀 Try this Example

View the complete source code on GitHub.

Steps to run:

1. Install fal:

pip install fal

2. Authenticate (if not already done):

fal auth login

3. Clone the demos repo and install dependencies:

git clone https://github.com/fal-ai-community/fal-demos.git
cd fal-demos
pip install -e .

4. Run the trainer:

fal run fal_demos/video/wan_trainer.py::WanLoRATrainerDemo

This demo uses a GPU-heavy training stack and can run for a while depending on your step count.

Deploy to fal

To host the trainer as an API:

fal deploy fal_demos/video/wan_trainer.py::WanLoRATrainerDemo

How it works

• Downloads WAN weights from Hugging Face and a pinned training repo.
• Prepares a dataset from your ZIP file (images and/or short videos).
• Runs LoRA training with DeepSpeed and returns the adapter weights.

Input format

Your training_data_url should point to a ZIP containing:

• Images or videos (.png, .jpg, .jpeg, .gif, .mp4)
• Optional caption files (my_clip.txt next to my_clip.mp4)

If you set auto_scale_input=true, videos are fit to 81 frames at 16 fps for WAN training.

Example request

import fal_client

result = fal_client.submit(
    "your-username/wan-lora-trainer-demo",
    arguments={
        "training_data_url": "https://your-bucket/train.zip",
        "rank": 16,
        "number_of_steps": 400,
        "learning_rate": 2e-4,
        "trigger_phrase": "your-style",
        "auto_scale_input": True,
        "video_clip_mode": "single_beginning",
    },
).get()

print(result["lora_file"]["url"])