Foreword

This model is the result of what I learned how LoRA models can be trained. This is still preliminary driven by the fact that there are a lot of parameters and settings which can be changed.

What the Model Does

Creates a glass ball with or without content.

Prompt

A good result can be achieved, when the topic comprises three parts like

blue ocean, island in a glass ball, blue cloudy sky

First and last part can sometimes be exchanged. One has to try this out. And one should experiment with something like this

island in a glass ball

glass ball with a island  

Exchanging the words in main topic can help to get what one wants.

Trigger Word

I used for the training a Trigger Word. But so far I tested this Trigger word is not necessary. The Trigger Word is

GlassBallConcept

Weights

A weighting under a value over 0.9 should be avoided. Using a weight of 1.0 one gets something that looks similar to the images with one blank glass ball without background I used for the training. I trained the LoRA model using images without background. A weight of 1.0 and above leads to deformed, distorted images with undefinable background.

Limitations

It is not all the time possible to to realize a desired theme.

Training

I used a limited set of 9 images for the training. A captions file was also used. By use of the possibilities of AUTOMATIC1111 9 images were resulting in at least 144 images for training. This was done to prevent the so-called overfitting.

The glass balls for the training were gray, light gray, light blue, white and pure white. The background was black, light gray or white.

To-Do

Improvement of this documentation. Refinement of the model. Test wise image creations.

Finally

Have a nice day. Have fun with the model. Be inspired!