how to add a new tag

[hyqyoung]

If I have several new pictures and want to create a new tag for these images, do you have any practical advice?

[altndrr]

If I understood the problem correctly, you would like to use CaSED to tag images, correct? If so, the quickest approach is to use the transformers implementation (see also Inference in the readme). Let me know if this responds to your question.

[hyqyoung]

If I understood the problem correctly, you would like to use CaSED to tag images, correct? If so, the quickest approach is to use the transformers implementation (see also Inference in the readme). Let me know if this responds to your question.

Apologize for not being clear enough. What I'm aiming to do is, without further training the model, have it recognize new labels based on just a few given images. This particularly involves new concept labels that the model has not encountered during pre-training, such as newly designed anime characters, among others.
In a more concise form: I want the model to recognize new, unseen labels (like new anime characters) from just a few images, without any additional training.

[altndrr]

I see two issues in this. First, suppose CLIP (the pre-trained model used as a base for CaSED) never saw some concepts in its pre-training. In that case, there is no guarantee it will be able to match them at inference (I guess this would be your case if, e.g., the anime characters you refer to were created approximately after 2021). Second, CaSED does not force any tag list, it simply uses semantic similarity to match images to web captions and later images to parts of these captions.

If you have access to a pre-defined list of tags you are interested in, you can try using CLIP directly. If performance is unsatisfactory (and training is not an option), you could consider an approach like TIP-Adapter, using the few annotated images you have to construct a cache to update the model predictions.

Let me know if you have other questions.