I just did fine-tuning and am now getting 'AssertionError: You do not have CLIP state dict!' Does an
I just did fine-tuning and am now getting 'AssertionError: You do not have CLIP state dict!' Does anybody know what could be the cause?
Fdo you have clip downloaded accurately?
Ffine tuned model is only unet
Mi am not sure. I think i will start over and see how it goes.
Fdid you use our onetrainer config?
Fyou can use kohya extract lora feature
Fi recommend dreambooth and extract lora
Fso dreambooth also bad quality?
Fok that is good
Fso all you need to do is extract lora and use on same base model
Fof course you can 
Fonetrainer generated model file same as kohya
Fyou have your safetensors right?
Flike 6 7 gb
Fyep
Fbecause onetrainer has EMA
Fit improves
Aso doctor
Ais it better to always tirain on dreambnooth
Aif i want lora?
Ainsted of lora tab
Ais it bnetter quality
Ado the annoying lines go away?
Ffor sdxl yes
Ffor flux i think native lora training better
Fextraction not working great
Fprobably extraction logic is broken not accurate
Ffor flux
Athen 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