Fine-tuning is a feature that offers a more personalized AI experience. This allows you to train a model to understand and recreate specific styles, faces, or objects.
PC: Find it in the main menu header.
Mobile: Open the dropdown menu by tapping on your profile picture (top right corner) and select “My Models”.
Choose your Model Type: Options include face, object or animal, and style. Name your model for further use.
Note: Fine-tuning is a PRO-only feature, but free users get 1 free face-model tune and 10 generations with it.
Choose images for model training:
Click on Choose or Create Dataset - here you can upload or choose images from your existing library that you want to use for training your model.
Start with at least 20. For more effective learning, diverse and numerous images are preferred. Once set, scroll down, agree to the terms, and click on "start training".
Note: A dataset is a set of images that are used to train a model.
Your models are private and safe. You are the sole user who can use them.
To contact NightCafe Staff directly, reach out directly via the feedback/support form.
"lora" or "LoRA" is the type of finetuned model that you can train on NightCafe. It is an acronym for "Low Rank Adaption" and is a method for quickly fine-tuning models on a small dataset. Simply said, the LoRA training model makes it easier to train Stable Diffusion on different concepts, such as characters or a specific style.
When you use a finetuned model, you need to add the token for that model to your prompt so that it can be interpreted in the context of the rest of the prompt.
The token is in the format <{type}:{name}:{optional weight}>. An example is <lora:My Face:0.8>. This allows you to write a prompt like "A photo of <lora:My Face:0.8> riding an elephant", or "A unicorn in the style of <lora:Dark Fantasy:0.5>".
The weight is optional and can be omitted. Weights should usually be between 0 and 1. If omitted it will default to 0.8. E.g. <lora:My Face> will be interpreted as <lora:My Face:0.8>. In the future there might be more types of models, which is why it's used as part of the token.