then why you say you reccomend it
then why you say you reccomend it
Ayeah only problem is the line artifacts that come up
Aswarm
Ai put 2k res
Athe lines only appear when using a lora
Ayeah
Ai found supir
Aso it gets the job done
Abut its incconvienti you know
Adb fixes those artifacts
Ajust amazing
Ai never tried to upscale in swarm
Acan you guide me?
Asure
Fi dont recommend for flux anymore
Fafter i did some more testing and usage
Fbut it has 1 advantage
Fif you upscale to 2x, extract lora works better than lora since lora produces lines
EHi guys! I've been training FLUX LoRAs using Kohya on Windows with 4 RTX GPUs. About a month ago, everything worked fine, but now I keep getting crashes related to torch.distributed.get_world_size() when using Accelerate.
Has something changed recently in PyTorch or Accelerate that breaks multi-GPU training on Windows?
Is there any recommended workaround for this?
Has something changed recently in PyTorch or Accelerate that breaks multi-GPU training on Windows?
Is there any recommended workaround for this?
Fyes you need like
FSXM or NVLINK machines
AHi Rex , when did he state that?
Aso whats the workaround
Asince lines dont go away, maybe train on a custom checkpoint?
Aon 2k res with loras?
Ahow
A1024x1024
Abut i use community loras too so idk theirs
Asuch as?
Acan you dm me an example dataset?
Aso for 1024x1024 i shouldnt go higher?
Awith dreambooth im not getting lines tho and it works perfect
Ayeah
Aand i tested in extreme resolutions
Aand results where outstanding
Athat's why your comment on not training the full model anymore confused me
Ai know its overkill for 20images to train the whole network but if it works it works
Auntil a solution arives
Aon replicate?
Aostris?
Aoh you dont use kohya
Ahow so
Ashould i train that ? what do you think
Awhat does wip stand for
Adb of course
Ain visual quality more
Ai trained a subject with 5 images
Avery limited dataset
Aand results where amazingf
Ai never extracted a lora
Ai want to try it at somepoint