Replace Blip2 with LLava (or keep both) for creating captions

[cwiggi01]

I have been looking at LLaVA woudl be nice if someday it could be integrated like BLIP2 has.

LLaVA: Large Language and Vision Assistant

https://github.com/haotian-liu/LLaVA#evaluation

More info:
https://arxiv.org/abs/2304.08485

Paper which shows its comparison to BLIP2
https://arxiv.org/pdf/2310.03744.pdf

[jhc13]

Thank you for the suggestion. I just tried it out and it seems pretty good, although I did notice some hallucination.

Unfortunately, LLaVA has not yet been added to the Transformers library (unlike BLIP-2), so it is difficult to integrate it into TagGUI.
However, I noticed work being done on this issue (here and here); I will consider adding LLaVA once this is complete.

[jhc13]

LLaVA has been added in v1.9.0.