Foreword

The current Embedding model was developed to create sitting mechanical cats of various kinds. However, as you can see in the gallery, the model is not limited to cats.

The Embedding model was trained using some of my photos with mechanical cats in a studio atmosphere. All of the cats in the photos are sitting. The background of all photos is misty.

What can the model do

The Embedding model is able to generate photos ranging from real to surreal, which are either photorealistic, pseudo-photorealistic or surreal-photorealistic. At least that's what it was developed for.

The model is preferably used for colored pictures, as shown in the gallery. More monochrome cats do not have a special effect. I like the colored more than the others.

How to use the model

The model works well with and without Hires.fix and Denoising Strength. The weighting works between 0.3 and 0.9. But the weighting is best between 0.4 and 0.8. Other values have to be tested.

One will need a little patience when using the Embedding model. My really first attempts with the Embedding model were more than unsatisfactory. If one then work a little bit on the underlying Prompt and play with the settings, impressive results are possible.

Advantage

A very big advantage of embedding is its size. Compared to checkpoint models, it is several orders of magnitude smaller. Compared to LoRa models and even Hypernetwork models, it is small.

Model Evolution

One can see an evolution in the quality of the images between the versions I create based on the same training images.

In contrast to the other models, I have changed the training a bit. This results in a forced sitting cat.

Model training

The Embedding model is trained on a local installation. The hardware is less than state of the art. Software is state of the art.

For the training of the Embedding model I used the model AbsoluteReality, which can be found here on the platform.

The pictures used for the training were all taken by myself. They show a cat sitting in front of a monochrome blurred background.

PickleTensor vs. SafeTensor

A training with the AI Web UI AUTOMATIC1111 results in the so-called PickleTensor. I have on my to-do list, that it would be better to prepare a SafeTensor. Up to now, I was not successful in converting PickleTensor to SafeTensor with reasonable effort. This is on my agenda when I have a little bit more time.

Finally

Have fun! Be inspired!

Donation

About each Buzz I am happy! Yellow ones are very welcome!

I am also happy to accept donations in one of 250+ crypto currencies:

https://nowpayments.io/donation/zentrocdot

Leave optional a comment like improve documentation or improve model etc. when doing so, that I know in what you are interested in.