>>/dobrochan/1220@603
I know how programmers like to let hype amateurs.

This is the first time I've seen so many people say, "I hate hype, I'm tired of the neural networks from the smoothies, but it's different, I'm scared, it can do what I get paid for."

All right, whatever.

Little news: quietly released model OpenAI whisper v2. She is noticeably better than the previous one and makes perhaps fewer mistakes than the average stenographer. He even knows where to put quotes, for example. Google and Apple voice input, dragon dictation and other alternatives were not nearby.

It works many times faster than realtime on a weak GPU, if you put https from here: / /github.com/ggerganov/whisper.cpp

>>/dobrochan/1221@603
I don't know what you want to do. In fact, this is a way to either understand which keywords give a given picture, or what pictures are obtained for such keywords. You probably want the first one. But your pictures are either human-made or generated by something like AnythingV3. Lexika searches only from the database of downloaded generation base stable diffusion. I think you'll have to wait until the Wifu Labs release their clip Interrogator, because it'll be able to pull the danburu tags out of the pictures (hopefully). While you try.

https://huggingface.co/spaces/pharma/CLIP-Interrogator
Or better yet, locally.
https:/ /github.com/KichangKim/DeepDanbooru