What does the 'TOK' mean?
Is this a useful word to learn some styles? Hope for explanation!
Going off the documentation in HF's SDXL lora training colab it's a sort of identifier:
Now let's add the concept token identifier (e.g. TOK) to each caption using a caption prefix. Feel free to change the prefix according to the concept you're training on!
- for this example we can use "a photo of TOK," other options include:
- For styles - "In the style of TOK"
- For faces - "photo of a TOK person" - You can add additional identifiers to the prefix that can help steer the model in the right direction.
- e.g. for this example, instead of "a photo of TOK" we can use "a photo of TOK dog" / "a photo of TOK corgi dog"
And in this arxiv paper they state it's a unique identifier for the trained style
We utilized a subset of 50 of the prompts used for the
dataset curation of the SDXL and PixArt-α model
comparison and used the same random seeds across
models. Prompts were modified for input into SDXL with
“in the style of Monet” appended to the end. Similarly,
prompts were modified for the LoRA model to append “in
the style of TOK”, where “TOK” is the unique token
string that was used during training to refer to the concept
in the input images, the style of Monet’s paintings.
Are we to replace TOK with something else such as Disney or do we leave TOK?