CatOnMars

Foreword

I built this Embedding model to study some effects while training this Embedding model. In this way the expectations of the results should not be set too high.

One big advantage of Embeddings and Textual Inversions are the file size in download and in local storage as well as the flexible usage. On the other hand side it is quite fast to train such a model relatively seen.

What the Model does

The embed turns a cat into a Mars style cat. Using keywords in the prompt such as "cat" and "Mars" will work best.

The Embedding and the Checkpoint Models

A bunch of models will work when using the CatOnMars Embedding. The images from the gallery give an impression of this statement. But I have to say, I prefer some special models, based on the fact, that they give me fast results.

Model Training

I used 12 to 34 image of my own pictures for the training. The Embedding Learning Rate ranged from 0.05 to 0.0005. The maximal number of steps were 15000. 5 Prompts were used for the training. [name] and [filewords] were used in the Prompt template file. The later one is using the words from the image file name for the Prompt. 8 vectors per token were chosen.

Farewell Message

Have a nice day! Have fun! Be inspired!