Textual Inversion Techniques

A training technique that creates a new embedding (token) representing a specific concept, style, or object. The embedding can be invoked in prompts with a trigger word.

Textual inversion trains the model to understand new concepts by creating custom embeddings.

How It Works

  1. Provide 3-10 reference images
  2. Choose a trigger word (e.g., “my-style”)
  3. Train a new embedding
  4. Use trigger word in prompts

Embeddings vs LoRAs

AspectTextual InversionLoRA
File sizeTiny (5-50KB)Small (10-200MB)
Training timeFast (15-30min)Medium (30min-2hr)
FidelityMediumHigh
FlexibilityHighHigh
What’s trainedNew tokenModel weights

Common Uses

Style Embeddings

Capture a specific artistic style:

  • Color palettes
  • Brushwork patterns
  • Artistic techniques

Concept Embeddings

Define specific objects or elements:

  • Custom products
  • Specific patterns
  • Unique textures

Negative Embeddings

Improve quality by capturing what to avoid:

  • “bad-hands-5” - Common hand problems
  • “easynegative” - General quality issues

Using Embeddings

  1. Download .pt or .safetensors file
  2. Place in embeddings folder
  3. Use trigger word in prompt
  4. Can combine multiple embeddings

Training Tips

  • Use consistent, high-quality images
  • 5-7 images often sufficient
  • Training 3000-5000 steps typically enough
  • Test with variety of prompts