textual inversion

Textual inversion learns new text tokens representing specific concepts for diffusion model generation. **Approach**: Instead of fine-tuning model weights, learn new embedding vectors that can be referenced in prompts. Model stays frozen. **Process**: Images of concept → optimize new token embedding to reconstruct images when used in diffusion → embedding stored as small file (~few KB). **Example**: Learn "" token from cat photos → prompt " wearing a hat" generates that specific cat. **Technical details**: Only optimize embedding (768-1280 dimensional vector), freeze U-Net and text encoder, typically 3000-5000 training steps. **File size**: Extremely small (~3-5 KB per concept) vs LoRA (~4-100 MB) vs DreamBooth (GB). **Limitations**: Less expressive than weight fine-tuning, may struggle with complex concepts requiring model modification, works best for styles and simple objects. **Use cases**: Art styles, simple objects, textures, color schemes. **Combining concepts**: Multiple textual inversions can be used together in same prompt. **Comparison**: Most parameter-efficient but lowest fidelity; LoRA is good middle ground; DreamBooth highest quality but most expensive. Choose based on quality vs efficiency needs.