i have one more question. Do you think models capability of following prompts decrease if we train t

Ddi_en_ey i have one more question. Do you think models capability of following prompts de...

F

Furkan Gözükara SECourses•9/16/24, 1:23 PM

i didnt notice such thing

A

anikey•9/16/24, 1:46 PM

is there any place where i can train a flux lora for free ?

A

anikey•9/16/24, 1:47 PM

i understand that i can make one for just 3$ on civit ai , but the thing is i am underage so i do not have access to online payment thing hence looking for a free alternative

Z

Zono50•9/16/24, 1:49 PM

@Dr. Furkan Gözükara I see in the supir app you can upload a dreambooth model for face correction, can we use a LORA instead for low vram?

Aanikey i understand that i can make one for just 3$ on civit ai , but the thing is i am...

F

Furkan Gözükara SECourses•9/16/24, 2:02 PM

no platform for free

F

Furkan Gözükara SECourses•9/16/24, 2:02 PM

it requires bf16

F

Furkan Gözükara SECourses•9/16/24, 2:02 PM

free platofrms dont provide such gpu

ZZono50 @Dr. Furkan Gözükara I see in the supir app you can upload a dreambooth model f...

F

Furkan Gözükara SECourses•9/16/24, 2:02 PM

nope

F

Furkan Gözükara SECourses•9/16/24, 2:02 PM

it uses a base model

Z

Zono50•9/16/24, 2:08 PM

ah ok any recommendations for improving images of self on those trained on lora?

Z

Zono50•9/16/24, 2:08 PM

using SUPIR?

V

Vinmi•9/16/24, 2:43 PM

@Dr. Furkan Gözükara how big (mb wise) is the LoRA you did with the 15 image dataset?

ZZono50 using SUPIR?

F

Furkan Gözükara SECourses•9/16/24, 3:01 PM

yep

VVinmi @Dr. Furkan Gözükara how big (mb wise) is the LoRA you did with the 15 image dat...

F

Furkan Gözükara SECourses•9/16/24, 3:01 PM

it depends several factors

F

Furkan Gözükara SECourses•9/16/24, 3:01 PM

but starting from 2 gb to 3.5 gb

F

Furkan Gözükara SECourses•9/16/24, 3:01 PM

single layer training arrived to kohya

K

Kallamamran•9/16/24, 3:05 PM

Kohya gets stuck here

K

Kallamamran•9/16/24, 3:05 PM

No errors

K

Kallamamran•9/16/24, 3:06 PM

Well I find this scrolling up

FFurkan Gözükara SECourses it depends several factors

V

Vinmi•9/16/24, 3:07 PM

I trained a 200 epoch LoRA (81 images dataset), with the following settings:

V

Vinmi•9/16/24, 3:07 PM

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

V

Vinmi•9/16/24, 3:08 PM

Got a 18mb LoRA out of it

V

Vinmi•9/16/24, 3:08 PM

It works, but is not incredible

VVinmi I trained a 200 epoch LoRA (81 images dataset), with the following settings:

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

i explained in video the factors

KKallamamran Well I find this scrolling up

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

set accelerator and gpu id 0

KKallamamran Kohya gets stuck here

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

it is using cpu that is the error

K

Kallamamran•9/16/24, 3:14 PM

Oh shit

K

Kallamamran•9/16/24, 3:16 PM

May I ask where? command line?

F

Furkan Gözükara SECourses•9/16/24, 3:17 PM

read

K

Kallamamran•9/16/24, 3:17 PM

KKallamamran Click to see attachment

F

Furkan Gözükara SECourses•9/16/24, 3:18 PM

ok then set accelerator

F

Furkan Gözükara SECourses•9/16/24, 3:18 PM

you have multiple gpus?

K

Kallamamran•9/16/24, 3:18 PM

I do

K

Kallamamran•9/16/24, 3:18 PM

I shitty internal GPU that gets listed

K

Kallamamran•9/16/24, 3:19 PM

Aha

K

Kallamamran•9/16/24, 3:19 PM

K

Kallamamran•9/16/24, 3:19 PM

accelerator device: cpu :/

K

Kallamamran•9/16/24, 3:19 PM

Well have to eat. Hopefully I can solve this later

K

Kallamamran•9/16/24, 3:19 PM

Thanks!

F

Furkan Gözükara SECourses•9/16/24, 3:20 PM

ok try gpu id 1

K

Koudkunstje (Roelof)•9/16/24, 3:57 PM

To test a Lora which guidance value is best? Like 1 or 3.5 or even higher?

K

Koudkunstje (Roelof)•9/16/24, 4:08 PM

I have a pretty hard time to find the sweet spot. Sometimes it changes the image quite a bit like on epoch 200 strenght 1.0 and on epoch 150 at strenght 1 .1

K

Koudkunstje (Roelof)•9/16/24, 4:22 PM

ok this image isn't very useful as discord resizes it

KKoudkunstje (Roelof)To test a Lora which guidance value is best? Like 1 or 3.5 or even higher?

F

Furkan Gözükara SECourses•9/16/24, 4:23 PM

cfg 1

F

Furkan Gözükara SECourses•9/16/24, 4:23 PM

flux guidance 4

F

Furkan Gözükara SECourses•9/16/24, 4:24 PM

i got a better config for fine tuning

F

Furkan Gözükara SECourses•9/16/24, 4:24 PM

it is nice

K

Koudkunstje (Roelof)•9/16/24, 4:27 PM

default shift (base 0.5 / max 1.15) ?

KKoudkunstje (Roelof)default shift (base 0.5 / max 1.15) ?

F

Furkan Gözükara SECourses•9/16/24, 4:35 PM

ah from your image i dont know what are they

K

Koudkunstje (Roelof)•9/16/24, 4:39 PM

if discord doesn't resize the image maybe this is more useful. Seem pretty spot on with epoch 200 and lora_strenght 1.0 at 1.1 if starts mutilate hands.

i have one more question. Do you think models capability of following prompts decrease if we train t

Similar Threads