Furkan Gözükara
patreon
JoyCaption is amazing to caption training data. Advanced Gradio APP with batch processing and 1-click install
🕑 Added 2024-08-24 02:13:07 +0000 UTCHere a Hugging Face space that you can test it yourself : https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha - still working
I have been requested to make a Gradio app for this so i made an advanced app and 1-click installers
It uses a clip siglip-so400m-patch14-384 and Meta-Llama-3.1-8B-Instruct as model and a fine tuned checkpoint for better captioning
My app who wants to checkout : https://www.patreon.com/posts/110613301
It has batch folder captioning feature as well and auto save all captioned images and captions into outputs folder
Also I have a very lightweight, super fast Gradio caption editor. Since I don't like other existing apps, i self developed this one from scratch : https://www.patreon.com/posts/108992085
Comments
Furkan Gözükara
well i think you need further fine tuning. sadly i dont know how to do
Diggy Dre
I know that, I am asking that, for those of us that still want to train some SDXL stuff, is it possible to use this and automate what this makes into something that removes the redundant words (for SDXL) and leaves the relevant words as tags in a WD14 style list of keywords? I feel like the level of detail as far as the specific tags this could generate would be awesome for that if there was a way to automate it. So basically, backward engineering what this is capable of for those who still want to train SDXL.
Furkan Gözükara
WD14 is very specific and works only on specific fine tuned sdxl sd 1.5 models. it is pretty worthless for flux
Diggy Dre
Is it possible to get this level of precision into a WD14 style of caption list for training SDXL? This is far more superior than WD14 in it's description.
Furkan Gözükara
this alone uses some real vram :D but would be good addition like auto load and deload
Neo Geo
Would love for you to embed this into SUPIR!
