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
- Provide 3-10 reference images
- Choose a trigger word (e.g., “my-style”)
- Train a new embedding
- Use trigger word in prompts
Embeddings vs LoRAs
| Aspect | Textual Inversion | LoRA |
|---|---|---|
| File size | Tiny (5-50KB) | Small (10-200MB) |
| Training time | Fast (15-30min) | Medium (30min-2hr) |
| Fidelity | Medium | High |
| Flexibility | High | High |
| What’s trained | New token | Model 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
- Download .pt or .safetensors file
- Place in embeddings folder
- Use trigger word in prompt
- 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