@Dr. Furkan Gözükara koha training is running much faster with the new version
@Dr. Furkan Gözükara koha training is running much faster with the new version
Dthere is some news about this?
LOne user wrote to me that my workout has problems due to the weight of 1gb. For some reason, his models are 18 mb each. I have seen such models with a similar weight, but they did not bring me high-quality results and differences
Fsingle layer
Freduces quality
Fthey dont test anything properly to compare
Fwhich version? torch 2.5?
Fyou can use lora resize of kohya and see quality difference
Lthanks
Z@Dr. Furkan Gözükara on supir app, if guidance scale needs to be 2, where do i set that? Text guidance scale?
DI'm not sure, I updated to train individual layers but with not a good success, now I training all layers and is super fast
D
Dresolution? bs? card?
D
D1024x1024
D
D
D1024x1024 rtx 3090
D
Dthis is normal?
Dno, for 1024 that's fast
Dthe checkpoint is 1.18gb so I'm sure I'm training all the layers
Dwhat did you have beefore the pytorch upgrade?
Dfalse alarm sorry
Dmy dataset es low ress
Dthats why
DI have bukets enable
Dah yeah, no upscale on kohya. resolution makes a major difference on flux
Dwhat resolution was it actually?
Dsome 768 some 512
Dnot all 1:1 diferent aspect ratios
MThanks. I posted there but there is a long conversation already happening so probably missed. I'll try posting again later
Fgood speed
Fnope
Fyep i was gonna tell this :d
TNew supporter here. Wanted to say thanks. Your info is great. Couple of questions for you. I'm doing LoRA training on Flux Really, really slow on my 3090 (Windows with latest torch 2.5 speedup), 51 images in LoRA. This may have been answered somewhere, but previously with this SDXL LoRA I would do 32 repeats, 51 images, 10 epoch (16320 steps). This was a little high admittedly. The config you had would be 10200 steps Are epochs better than repeats with FLUX1.dev? Does this lower the number required? Just looking to understand that a little better. As for the slowness I'm 4735 steps and running on 1.5-2 days. I did update torch. Pretty much the same speed for me. Might have to try on RunPod. Any thoughts/advice are appreciated.
Fwhat is your step speed?
Fyou can use Rank_4_17250MB_4_85_Second_IT.json it is pretty decent speed
F200 epoch may be a lot for 51 image
Fbut you can compare checkpoints
XIs nsfw possible with flux lora?
Fi dont know i really dont into nsfw
Fwe also avoid discussing it here
XOkay! Asking for someone else
DI've just trained a very bad dataset on flux. Blurry, mixed resolution down to 128px. For SDXL, I have carefully prepared it using SUPIR and more. With flux, it just works without preparation. This model is insane
Fwith flux you can only get better 
Fi plan to do general fine tuning with our 20k images but i need at least 2x A100 like 1 week
F48gb fine tuning fails at multi gpu
