I may try the new captioner for a style training
I 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 
Ngetting this after trying to start. Issue with dir name fixed, new ones come))
Ntried to start rank_5
Fuse dev fp16 model
NI'm using this one
Fyes it is correct one
Fwhat is your gpu model?
Nmaybe t5 is wrong?
Fyep it may be wrong :d
Nrtx 4060ti 16gb
Fuse the downloader bat file i have
Nhad to extract only t5 from your python, because downloading flux again will piss me off)
Nthx
Fyou can edit my files
Fremove the ones you dont need
Fall my files are editable easy
Nyep, just did that, thanks
X
Xjust put
kohya_ss\wd14_tagger_model, didnt work
Xnvm, downloaded another model from kohya and replaced that model with eva02, it works
X
