WAN 2.5 Video Finetune

What is WAN 2.5 Video Finetune?

WAN 2.5 Video Finetune allows you to train custom video generation models using your own video data. The trained model can generate videos in your unique style, preserving the specific visual characteristics from your training dataset.

How should I prepare my training dataset?

Create a ZIP file containing 5-20 training videos (MP4 format, 3-10 seconds each) and a dataset.jsonl file. Each line in the JSONL file should have format: {"video": "filename.mp4", "text": "description of the video"}. Higher quality, consistent videos lead to better results.

How long does training take?

Training typically takes 6-8 hours depending on your dataset size and configuration. You can monitor progress in real-time and will receive a notification when training is complete.

What are training epochs and learning rate?

Epochs determine how many times the model processes your entire dataset. More epochs generally improve quality but increase training time. Learning rate controls how quickly the model adapts—higher values train faster but may be less stable.

What is LoRA Rank?

LoRA Rank determines the dimension of the low-rank matrices used for adaptation. Higher ranks (32-128) capture more detail but require more memory and training time. Lower ranks (8-16) are faster but may capture less nuance.

How much does training cost?

Training costs are based on token consumption, calculated as: Total Tokens = Σ(video billing duration) × (max_pixels / 1024) × n_epochs. Current rate is $0.05 per 1000 tokens. Typical training costs range from $5-50 depending on dataset size and epochs.

Can I use my trained model commercially?

Yes, you fully own the models you train with WAN 2.5 Finetune. You can use generated videos for commercial purposes including marketing, advertising, product showcases, and content creation.

How do I deploy my trained model?

After training completes, select a checkpoint and click 'Deploy'. Deployment takes 5-10 minutes. Once deployed, your model is ready to generate unlimited videos from input images in your custom style.

What if training fails?

If training fails, you can check the error logs and retry with adjusted parameters. Common issues include insufficient training data, corrupt video files, or incompatible formats. Our support team can help troubleshoot specific issues.

Can I train multiple models?

Yes, you can train and deploy multiple custom models. Each model can have its own unique style based on different training datasets. Manage all your models from the dashboard.