LoRA Training & Usage

Create and use LoRA models for personalized AI generation.

What is LoRA?

LoRA (Low-Rank Adaptation) allows you to fine-tune AI models with your own images, creating personalized styles or consistent characters.

Why Use LoRA?

Benefits:

• No need to describe appearance in every prompt
• More consistent results than prompt-only generation
• Train once, use unlimited times

Training Your Own LoRA

Create personalized AI models without coding:

Step 1: Prepare Your Dataset

1. Collect 10-20 diverse images

   • High quality, no watermarks or text overlays
   • Multiple angles and poses of your subject
   • Varied lighting and backgrounds
   • Subject clearly visible in each image
   • Consistent image dimensions recommended

2. Compress into a .zip file

   • All images in one folder
   • Supported formats: JPG, PNG, WebP
   • Optionally include .txt files with captions (same filename as image)

Step 2: Configure Training

Step 3: Train

1. Go to a LoRA trainer model page (e.g., LoRA Trainer)
2. Upload your dataset
3. Configure parameters
4. Click Run and wait for completion

Browse all LoRA trainers at wavespeed.ai/models — filter by “Training” category.

Note: Training time depends on dataset size and parameters. Typical estimates: ~8 minutes for 1000 steps, ~25 minutes for 3000 steps. If a timeout or system error occurs, refunds are processed automatically.

Step 4: Use Your LoRA

Once training completes:

• Copy the URL provided to use directly in API calls
• Download the .safetensors file for local use
• Upload to Hugging Face for permanent storage and easy sharing

Using Existing LoRA Models

Apply LoRA models from Hugging Face or Civitai in your generations.

Where to Find LoRA Models

Requirements

API Example

curl --location --request POST 'https://api.wavespeed.ai/api/v3/wavespeed-ai/qwen-image/edit-plus-lora' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer ${WAVESPEED_API_KEY}' \
--data-raw '{
  "prompt": "Turn the camera to a close-up.",
  "images": [
    "https://example.com/your-image.png"
  ],
  "loras": [
    {
      "path": "dx8152/Qwen-Edit-2509-Multiple-angles",
      "scale": 1
    }
  ],
  "size": "1536*1536",
  "output_format": "jpeg",
  "seed": -1
}'

Note: Parameters vary by model. Check the model’s README for all available options.

LoRA Parameters

Tips

Trigger Word:

• The trigger word activates your LoRA’s trained style/subject
• Must include it in your prompt — without it, the LoRA may not take effect
• Example: If trigger word is “p3r5on”, your prompt should be “p3r5on walking in a garden”, not just “a person walking in a garden”

Scale:

• Controls how strongly the LoRA affects the output
• Too low (< 0.5): Effect barely visible
• Optimal (0.8-1.0): Balanced results
• Too high (> 1.5): Over-stylized, may cause distortion

Combining Multiple LoRAs

Some models support combining high-noise and low-noise LoRAs for more control:

curl --location --request POST 'https://api.wavespeed.ai/api/v3/wavespeed-ai/wan-2.2/i2v-720p-lora' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer ${WAVESPEED_API_KEY}' \
--data-raw '{
  "prompt": "A person walking through a garden",
  "image": "https://example.com/your-image.jpg",
  "duration": 5,
  "high_noise_loras": [
    {
      "path": "owner/high-noise-lora",
      "scale": 1
    }
  ],
  "low_noise_loras": [
    {
      "path": "owner/low-noise-lora-1",
      "scale": 1
    },
    {
      "path": "owner/low-noise-lora-2",
      "scale": 1
    }
  ],
  "seed": -1
}'

Tip: When combining LoRAs, start with lower scale values and adjust gradually.

Compatibility

LoRA must match the base model exactly:

• Same model name
• Same version
• Same parameter size

For example, a LoRA trained on FLUX Dev won’t work with FLUX Schnell. A LoRA for Wan 2.1 14B won’t work on Wan 2.1 5B.

How to check compatibility:

• On Hugging Face: Check the Model Card for “base model” info
• On Civitai: Check the “Base Model” field in Details section

Troubleshooting

Learn More

• Train LoRA Guide
• Use LoRA Guide