Furkan Gözükara
patreon
Sota Image Captioning Model Kosmos-2 Added To Our Scripts Arsenal
🕑 Added 2024-02-15 01:42:40 +0000 UTCJoin discord and tell me your discord username to get a special rank : SECourses Discord
Our Scripts arsenal here : https://www.patreon.com/posts/90744385
We have captioners_clip_interrogator , LLaVA_auto_install , CogVLM , Qwen-VL , blip2_captioning , Kosmos-2
You can see full post here : scripts_arsenal_full_screenshot.png
Kosmos-2: Grounding Multimodal Large Language Models to the World > https://github.com/microsoft/unilm/tree/master/kosmos-2
- I have modified it and added batch processing too
- Download Kosmos-2_v1 .zip and extract into any folder you want to install
- All our scripts generate a new separate venv so they will never conflict with other apps
- Double click install_windows .bat and install
- It will install everything fully automatically
- Then use run_kosmos .bat file to start the app
Our Gradio APP both supports single image caption generation and also batch image caption generation as can be seen below
Comments
Furkan Gözükara
just added with V4 redownload :)
John Dopamine
This has been very helpful. If you ever make an update one thing I'd love added is an option to -skip- captioning any files that already have .txt files generated. I often run captioners on folders w/ many images, and sometimes may add new images after a set had already been captioned. If images w/ .TXT files already are skipped running the captioner is super quick and it'll just add captions for the few new ones. As it is now I have to put anything new in a seperate folder and run it on them / then move them otherwise it'll redo the whole set (which I may have edited to improve in somecases). Not a huge thing, but if you do revisit any of the captioners you've coded it'd be cool to have that type of option/toggle implemented.
Furkan Gözükara
yes LLaVA supposed to be strongest one. It also have the biggest model
Tech Meowpunk
Lava seems more accurate to me, I've tested a lot of bugs, maybe I didn't specify the accuracy controls
Furkan Gözükara
you are welcome.
Steve Bruno
This looks awesome thank you
