Pitfall Embedding

Foreword

One learns from mistakes. That is what I learned once again today. After preparing a LoRA model for lighthouses, I started creating Embeddings and Hypernetworks. The intention was, to compare the results from LoRA with the results of Embedding and Hypernetwork.

Problem Description

While working on Embeddings and Hypernetworks I made the stupid mistake, I am writing about. First I created an empty Embedding and an empty Hypernetwork. Then I tried LoRA again, and I got all the time photos of women instead of a lighthouse. What happened.

When on uses an Embedding in a Prompt one writes e.g.:

lighthouse, lighthouse:1.0

When on uses a Hypernetwork in a Prompt one writes e.g.:

lighthouse, <hypernet:lighthouse:1.0>

An experienced user will see, in which pitfall I was running, when using following Prompt e.g.:

lighthouse, <lora:LighthouseConceptLora:1.0>

The system is looking in the Embeddings directory by standard for an Embedding called "lighthouse.pt", which is there, but still empty. In other words one can say, the systems finds in the Embeddings directory a file called "lighthouse.pt" and is thinking, that this should be used in the Prompt as "lighthouse" trigger word. And so the problem took its course.

Troubleshooting

After realising that the naming was a mistake, I deleted "lighthouse.pt" in the Embeddings directory and restarted the system. Now, no wonder, I was again able to create lighthouse.

Conclusion

I learned to be careful in the naming of models, especially names for Embeddings. Hypernetworks and LoRAs have there special placeholder or trigger word, but Embeddings not.

Finally

Have a nice day. Have Fun. Be inspired!