Train Stable Diffusion on your own images to generate a specific person, style, or object consistently.

DreamBooth

• Fine-tunes the full model on 10–30 reference images
• Best for: a specific person's face, a product, a character
• Result: a full model that knows your subject
• Requires significant GPU and time (or use cloud services like Astria)

Textual Inversion (Embeddings)

• Trains a new text token to represent your concept
• Smaller file, less VRAM required
• Good for: styles, textures, objects
• Less powerful than DreamBooth for faces

LoRA Training (Best Balance)

Train a LoRA on your subject — smaller than DreamBooth, more powerful than embeddings. Use Kohya_ss trainer. The most popular approach today.

Reference:

DreamBooth training guide

https://huggingface.co/docs/diffusers/training/dreambooth