I have a RTX 3060 and am getting ~45sec/it, which would take quite long for 9200 steps, is that norm
I have a RTX 3060 and am getting ~45sec/it, which would take quite long for 9200 steps, is that normal?
Dnvitop is showing >90% GPU utilization
Fnice test
Fdreambooth tab
Frtx 3060 takes around 40 second / it in long run
DThe training was finished, resemblance is almost perfect, but it still looks a little undertrained, with regularization images it needs much more steps or increase the learning rate to get resemblance. I will resume the training from 160 repetitions per image and I will report back, there was no bleeding between the trained subjects
Fwow awesome
Fi am waiting that model to become more mature
Fwill fully test it and compare
DNice, If you have to train more than one subject it is your only choice, in my tests is superior on everything, the only problem is if you are going to do a full training the inference of the resulting model will be slower because you have to set the cfg scale to 3.5, to overcome it what can be done is to do the finetune, extract the lora and then use the resulting lora on regular flux-dev and be able to use cfg scale 1 in conjunction with distilled cfg
Fonce you use cfg > 1 = it doubles the duration of generation
Dthe quality of an extracted lora from a finetune is comparable to a finetune?
Fi made a comparison
Fbut i doubt it would work on
Fdev model
Dthe loras I'm training on flux-dev-de-distill works perfect on regular flux-dev, why the extracted lora will not work?
Dnot in the other way around the loras trained on regular flux-dev lost resemblance on flux-dev-de-distill on inference
Fwow if working then yes
Fshould work
Fhave you compared grid side by side?
Dno I just created some pictures to test
DI resumed the training from the last lora checkpoint
D@Dr. Furkan Gözükara on flux-dev-de-distill to get resemblance with regularization I used a much higher lr 0.004 on 160 epochs looks still a little undertrained, my lora rank and alpha is set like this, this can be a problem
Fyou basically changed the LR i researched
Freduced it like 1/8
Fi am offline 
Dohhhh, now I understands
DNo wonder it was taking me so long to train my LoRAs, the learning rate was too low 
D@Dr. Furkan Gözükara sorry that you are offline, no hurry to answer, in your best configs nowhere says if the config is a finetune or a lora
ELoRas are in Best_config folder inside kohya_gui_flux_installer.zip and finetunes are in Best_config_v5.zip
HIt's great to hear that! I just started training a LoRA about an hour ago, but I canceled it now because I'm going to rent 2 A100s for the training. Do you have any parameters that you recommend adjusting between De-Distilled and Standard flux LoRA training?
Dyou can use the same settings, the samples during training doesn't work, other than that the same, you can train multiple subjects in one lora with the de-distilled model and learn new concepts much better
Dthanks!!
HThank you so much 
DCan you resume/ increase the no. of epochs in the existing training? Or is it that you had paused it?
DIsn't 0.004 faster than 0.000004? Or is LR counter opposite in ML?
DSadly, my pose training has completely failed on both the models. Think I'm doing something wrong on the fundamental level itself 
, 'cuz I'm approaching it like a character fine-tune.
, 'cuz I'm approaching it like a character fine-tune.DGet what you're trying to say. Changing strategy, training different poses as multiple separate concepts instead of training them under the same poses class.
DInference on dev_distill can be pretty good as well, if you tweak it in Comfy. But yes, it is slower.
Dflux_dev versus flux_dev_de_distill
Dversus Pro 1.1
Dyou can get any lora checkpoint and resume the training from there, there is no pause option.
D0.0004 is much faster but it was my mistake, my lora rank was at 128 and my network alpha at 16 that means that I had to increase the lr up to 0.0004 to match the dr. Furkan settings, just use one of his lora configs and you are good to go. lora rank 128 alpha 1 or rank 128 alpha 128 is the same setting, because 128 divided 128 equals 1, all this iinfo is kohya's documentation
Dunless the poses you are training are very uncommon may be the pose is already trained on the base model, you have to prompt it correctly, if the pose is kind of unique then you can have a dataset of pictures with your especial pose and train it with captions like "nameofthepose pose"
DThat's how I'm captioning my images right now to start new training. But what about unique token and class token? Do I need to enter all of them comma seperated or I can just go into the kohya directory directly & modify folder names like 1_nameofthefirstpose, 1_nameofthesecondpose, and so on ?
DThank you! Would a 4060 be much faster?
Any other way to speed up?
Any other way to speed up?
Dyou can just change the folder name, you can use an unique token an space pose, so is in the class poses that the model already knows for example "pose1 pose" or name of the folder "1_pose1 pose" just a unique token without the class pose can confuse the model, that's my theory, I never trained a pose...