Fine-tune an image model

You can fine-tune an image model to generate images of yourself, or generate images of a particular style.

Here are three guides to get you started:

When should I use LoRA?

LoRA has a few differences from DreamBooth that make it especially appealing as an alternative:

  • Faster training: Training a new concept with LoRA takes just a few minutes.
  • Smaller outputs: Trained LoRA outputs are much smaller than DreamBooth outputs. This makes them easier to share, store, and re-use.
  • Multiple concepts: You can combine multiple trained concepts in a single image. (This feature is still experimental, but we’re working on improving it. 🧪)
  • Faster image generation: When you train your own DreamBooth model on Replicate, the model only stays warm when you’re actively using it. With LoRA, you’re not running your own model, but rather running the one cloneofsimo/lora model, which is always on and ready to serve predictions.
  • Better at styles, worse at faces. Based on our experimentation, LoRA seems to do a better job at styles than DreamBooth, but faces aren’t as good. They are stuck in uncanny valley, rather than looking precisely like the person. Your results might be better than ours though, so let us know how you get on.
Replicate’s GUI
People have been making some magical products with DreamBooth, such as PhotoAI and ProfilePicture.AI.