Hello again,batch 7 at 250 images on my 4090 is suuuuuper slow
Hello again,batch 7 at 250 images on my 4090 is suuuuuper slow
F
Fwas able to spin a 250 last night
Fwhat am I missing?
Fmessage.txt14.16KB
FUsing the 24G config file for dreambooth training
Ftrying to train a person
F
Fany help super appreciated
Fbatch size 7 is for 48 gb gpus
Fyou should use batch size 1 config
Ffor 24 gb gpus
F24GB_GPU_23150MB_10.2_second_it_Tier_1.json
FOh I see
Fthank you
Fby the way kohya made update
Fhopefully it will become even faster
Fi havent tried yet
FWill try it on the next one
Fit is not merged into main branch yet
FOh, OK, thanks
1so.. I've been training on massed_compute recently (A6000) but thought I'd give my 4090 a test. I'm actually quite suprised at the performance using the 24GB_GPU_23150MB_10.2_second_it_Tier_1.json
1fresh install of Koyha.
Frtx 4090 will be faster than a6000 with newest update - not merged into main yet :d
Ffor fine tuning
AWhat is the fastest kohya config to use if you have 80GB vram? (H100)
Fbatch size 7 flux fine tuning
Ydoesnt batch size decrese quality ?
Fit does
Fbut for speed mandatory
Swhen I try to train with dreambooth, the safetensor files are massive (23gb each). Is that to be expected?
Fyou can later convert into fp8 - half size
SI tried training one on a particular person, and the sample image during training looked perfect, but when I actually use it with flux, it just makes generic people. Is there something else I might've forgotten to do?
Fyep obviously :d
Fi dont know what though
SUpon finishing the training, can I simply move the safetensor files to the lora folder, and generate an image with them using the trigger word? or are there steps I'm unwittingly skipping?
Fthat is not lora
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fi assume you are telling 23gb files
FYouTubeSECourses
Ultimate Kohya GUI FLUX LoRA training tutorial. This tutorial is product of non-stop 9 days research and training. I have trained over 73 FLUX LoRA models and analyzed all to prepare this tutorial video. The research still going on and hopefully the results will be significantly improved and latest configs and findings will be shared. Please wat...
Slol, I understand. Thanks.
Z
Zand are the faster speeds in the current branch (since it's sd-scripts that are faster)?
V
Fi didnt test i copy pasted sd scripts to test yesterday
Fcoıuldnt start yet planning very soon hopefully
Zwhat did you copy, just sd-scripts flux branch?
Z@Furkan Gözükara SECourses Did you ever do a rank test on Lora Extraction? We know 64 is max, but 63 is only 1.2GB, is there really a difference? 64=12GB 63=1.2GB
ZPersonally, I have only done a small test, but with the wrong model, so it was hard to tell. I am training a new one at the moment, and will test again
WThe max rank that can be extracted is 640 not 64 my friend. Even training with a6000 we already achieve rank 128
