Embeddings is what you actually feed to the AI models , its a buncha numbers

Normally the tokenizer converts the prompt text to an embedding

But its possible to keep embeddings as .safetensor files . Being numbers , they do not have to correspond to written text

One can train by adjusting an embedding to create specific images for a given AI model. This is called textual inversion

Embeddings are not used alot with the natural language models , but were popular for SD1.5 and SDXL / Illustrious models

If the image output is y ,

the the checkpoint is like a square matrix A , multiply a matrix with a vector x , and you get another vector y , which is the output

Ax = y

The embedding is the vector x you multiply the matrix with.

The LoRa , is a matrix delta_A which you add to the matrix A in the model

so with LoRa , its

(A + delta_A) * x = y

//---//

LoRas are more popular , as the prompt x can be anything you want it to be

//---//

With embeddings , one can combine it into a written prompt , so if the keyword to trigger the embedding is 'beckham' , them I can write in prompt 'a beckham on a football field'

The 'beckham' is no longer a single token word , but is replaced by the 8x716 embedding , or however large it is

Embeddings are usually 8-16 token positions long

//---//

In summary , you can use LoRa and embeddings in combination . They act separately on the output