🔧Good SD1.5 default settings

Guidelines to get you started with the art of finetuning!

What are good default settings to train a model?

Training images

The first step is to assemble a good training set. These recommendations are valid no matter if you want to train a model on a person or an object.

  • Use at least 10 to 15 images for best results.

  • The images should be good: in focus, well lit, clearly showing the subject, without other major elements in the image.

  • The images should be varied: include multiple camera angles, lightning angles and backgrounds. When training on a person, it's important that they show various face expressions. Avoid accessories that cover the visage or excessive makeup, except if they're an integral part of the person's identity.

  • Adding more images will help, but only if the images are good. Adding mediocre images will decrease quality rather than improving it.

Training parameters

  • Finetuning steps: a good starting point is to 100 training steps per training image. So for example for 13 images, you would train for 1'300 steps.

  • Learning rate: don't touch it unless you really know what you're doing. The default is 1e-6 (0.000001)

  • Instance prompt: photo of ukj person is a good default when training on a person. Use photo of ukj object for an object, ukj style for a style.

  • Base model: stable-diffusion-v1-5 is the standard base model, powerful and versatile. realistic-vision-v1-3 is great for realism, and a good choice when training on people.

  • Image cropping: use "Face Crop" if you're training on a single person's face, otherwise leave it to "Center Crop".

  • Checkpoint format: leave it on "CompVis", the format compatible with the AUTOMATIC1111 web UI.

  • Use safetensors: leave it on, it saves the model in a more recent file format that ensures safety.

  • Callback URL: leave it blank.

  • Extract LoRa: turn it on if you intend to use a LoRa file.

  • Offset Noise: leave it off if you don't want to create images that are specifically very bright or very dark.

What about LoRa files?

The recommendations above are also good to train LoRa files. LoRa files are 20x smaller than full models (100MB vs 2GB), but may not reach the same level of quality.

Going further

This article from HuggingFace provides a good analysis in more details on how to train models: https://huggingface.co/blog/dreambooth

Ultimately, you will need to learn what works best for your specific subject and your specific images through trial and error:

  • Try out various number of training steps, training between 90 and 110 steps per images

  • Try out various base models

  • Try out various selections of training images

Last updated