Foreword

This Embedding follows the Base Model, LoRA and Hypernetwork, which I created with the topic red-white lighthouse, to study what each of the different models can do.

This model can have a positive effect on a picture, but it does not have to, according to my current experiments. Expectations should not be set too high.

What the Embedding Model Can Do

The Embedding model can modify a lighthouse in a way, that the lighthouse will be round with red-white colored sections.

Limitation

Embedding usually only creates a lighthouse without a detailed environment. It depends on what was brought into the training.

Base Model versus Embedding

The Prompts I am creating so good, that it is not possible to get more out of a photo with such an Embedding like this. So far, my attempts with embedding have been rather worse than the result without embedding.

Training

I trained the model, as I did before, with 12 lighthouse photos from my photo collection. The square images have a resolution of 512 x 512 pixel. The main content in each photo is the lighthouse.

In the Textual Inversion template text file I used [name] as well as [filewords]. For this each filename is a description what to see in the photo.

The used latent sampling method was random. The Embedding learning rate was set to 0.05:500, 0.005 to start fast and slow down later. The sample images show, that the first set of images can be far away from the trained topic.

Remarks

This is the worst functioning model I have ever created. I was not convinced that this approach to embedding would work and I seem to have been right. Embeddings are probably better suited for people or animals.

To-Do

Right now the computer is creating a new Embedding for testing purposes. I am using a reduced set on images. In addition to this I changed some other parameter for the training.

The interim results with the new embedding look promising 😀. Some preliminary results:

https://civarchive.com/posts/6798698

Finally

Have a nice day. Have fun. Be inspired!