There’s a way to get the joycaption to work faster? I’m getting 17s/image on a 4070 super
There’s a way to get the joycaption to work faster? I’m getting 17s/image on a 4070 super
Fnot really. on rtx 3090 it is like 9.5 second / it
Fbut you can rent multiple gpu and batch caption there :d
JIt might work better for style, cuz it really describes the style better than Florence2 SD3 Caption, but, I think, if you are dealing with character or a concept, Florence2 is WAAY faster, like, I captioned 70 images with it in like 3 minutes
Fye could be faster
Fyou can rent 10x gpu on runpod and quickly caption 
F
JWait, I changed the max token count and it’s much faster
JI don’t need a 300 token caption I think
Slightweight 
Fyep
Show do you add in a special token in joycaption for what I want to call the Lora? like ohwx etc?
Fuse our caption editor
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Sis that a separate program?
Fyep
Fyou can replace words append etc
Freally useful
HHave you tested training results with these super detailed captions vs just ohwx man?
Fi am about to share on patreon
HI noticed for my Lora’s where I just do a short caption it does affect the prompting of flux a good bit
HI may try the new captioner for a style training
Fadded to patreon post
Fcheck grids out
SDr Furkan i hope everything is ok, I received the patreon emial, but there is a step by step guide?
M
Fit is in record
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Fread top to bottom
N@Furkan Gözükara SECourses when you say you did some training on 1536x1536. this means you resized your training images to this resolution, yes? did you also modify "max_resolution": "1024,1024", in the config, or just let kohya do bucketing?
Fimages were 1536 and training config was 1536
Fi didnt upscale images they were raw resolution - downscaled from raw
Nnice - 90% of mine can be downscaled from raw to 1536 so i was going to try... increasing max resolution to 1536 and applying t5 attention max will still fit in 24GB?
MNall of his tests so far concluded it made the resemblance on character lora's worse than not using
Fi tested them
Fdoesnt help
Flet me put to top of post
F
Fit should fit
Fbut it will be slower
Donly square pics to train with or for example sdxl resolutions work too? like 1152x896, 1216x832
Dfrom what I read Flux is really flexible with training
Nnot quite - abort, abort! lol
Fit should work very elastic
Flooks like step 2 not installed
Fhow much vram you use before starting train
Fah 1536 didnt fit
Ni had 1536 and t5 mask checked... i unchecked t5 and left the 1536, running using 21.2GB vram but getting about 26s/it 