Prior Prediction

Prior Preservation Background:

A long term project by dxqbYD to add a form of model preservation to Lora Training. Originally developed for use with Flux, as that model tends to more easily see lost of concepts when training a Lora.

Usage

In order to use prior prediction, you need to do the following:

1. Create a new concept, that will be used as the prior prediction
2. In the new concept, choose prior prediction as the concept type.

That is it. Now you have a concept that will be used as prior prediction for your training run.

Limitations

• Only works for Lora training.
• Was primarily developed for use with Flux

Guidance

• What you should you NOT use for your prior prediction concept?

  • Your training dataset with the same captions. The model will be confused by this setup, so do not do it.

• What should you use for your prior prediction set?

  • Related concept with simple captions
  • Related concept with no captions

• What should the ratio of images be?

Technical Information

What does prior prediction do at a technical level?

If a concept is marked as a prior prediction concept, the training target is changed. Usually the model is trained to denoise any image into the training image. For prior prediction images, the model is trained to denoise the image into the same thing the original model would. It's essentially trying to prevent the model from changing for those training images