Software Engineering Courses (SECourses)•3y ago

One thing i loved with lora in khoya is it train fast like 10-15mins and ur done. Unlike dreambooth

One thing i loved with lora in khoya is it train fast like 10-15mins and ur done. Unlike dreambooth takes 1-2hrs depend on ur data and settings. I hope Lora in khoya will dev better training settings. I cant produce precise face from my subject. Annoying

justalurker•3/8/23, 5:20 AM

Ah, finally got Dreambooth working in the Colab.

Jjustalurker Ah, finally got Dreambooth working in the Colab.

TrimacOP•3/8/23, 5:23 AM

Was it Smooth there in colab?

TTrimac Was it Smooth there in colab?

justalurker•3/8/23, 5:30 AM

Well, despite having spent my whole career (I am now retired) as a Systems and Network Engineer I have almost no coding ability, so it could have gone smoother.

justalurker•3/8/23, 5:37 AM

Here was a pretty good output. Took about 15 years off my age, so I approve.

P1pSl1p•3/8/23, 5:38 AM

at 1.8 im getting "looks like" results. but I'm going for a 1:1

P1pSl1p•3/8/23, 5:38 AM

Kohya feels like textual inversion with extra steps and HD usage

P1pSl1p•3/8/23, 5:39 AM

I got a lot of "looks like" with textual inversion. not until I switched to dreambooth did I start getting 1:1

Original message was deleted

Steel•3/8/23, 5:50 AM

what are your parameters?

P1pSl1p•3/8/23, 5:50 AM

exactly the same as the video

Steel•3/8/23, 5:51 AM

the video was only for installation

Steel•3/8/23, 5:51 AM

if you used 512 photos you failed

P1pSl1p•3/8/23, 5:51 AM

he ran through a trial run as well

Steel•3/8/23, 5:52 AM

show me your settings

Steel•3/8/23, 5:52 AM

also what base model

P1pSl1p•3/8/23, 5:53 AM

closed now. I'm going through my dataset.

TTrimac He only teach how to run the lora in khoya but didn't get the main goal of produ...

Steel•3/8/23, 5:54 AM

I said not to listen to tutorials, the link was only for installing kohya

P1pSl1p•3/8/23, 5:55 AM

let's move this to DM so we don't clog up the main chat?

TTrimac One thing i loved with lora in khoya is it train fast like 10-15mins and ur done...

Steel•3/8/23, 5:55 AM

if your training takes 15 your dataset is too small

PP1pSl1p let's move this to DM so we don't clog up the main chat?

Steel•3/8/23, 5:56 AM

SSteel if your training takes 15 your dataset is too small

TrimacOP•3/8/23, 7:00 AM

Interesting

TrimacOP•3/8/23, 7:05 AM

On going training on dreambooth. I wanted to cancel my training progress since im seeing better result of my subject in preview image. However, I forgot to check "Generate ckpt file when training is cancelled" but "Genenrate ckpt file when saving during training". Was checked. Im a good to stop the training now?

P1pSl1p•3/8/23, 7:08 AM

check your stable diffusion models folder

P1pSl1p•3/8/23, 7:09 AM

you should see a folder in there named your training model

P1pSl1p•3/8/23, 7:09 AM

in there you should have a list of ckp files that correspond with the training steps

P1pSl1p•3/8/23, 7:10 AM

if you have that then you should be good to stop the training and test using those checkpoints

PP1pSl1p if you have that then you should be good to stop the training and test using tho...

TrimacOP•3/8/23, 7:11 AM

Oh! i forgot lol! Thank for saving my time. Let me check

Aurora 神名河•3/8/23, 7:29 AM

Hi I have got some strange problems.

Aurora 神名河•3/8/23, 7:29 AM

This time it is the multiple controlnet

Aurora 神名河•3/8/23, 7:29 AM

I can't use use it for some reasons

Aurora 神名河•3/8/23, 7:30 AM

I can only use it for only 1 controlnet per generation.

Aurora 神名河•3/8/23, 7:31 AM

If I tried to use 2 controlnet in the same time only the first controlnet will be applied.

P1pSl1p•3/8/23, 7:53 AM

I take back what I said about Kohya Lora training. @Steel has shown me the light. Thanks mate, incredible results.

Aurora 神名河•3/8/23, 8:12 AM

What is your training setting sir? Could you share it here?

PP1pSl1p I take back what I said about Kohya Lora training. @Steel has shown me the light...

TrimacOP•3/8/23, 8:12 AM

How many images and Learning rate you used? loooks good tho.

P1pSl1p•3/8/23, 8:14 AM

77 images. not cropped. full resolution

P1pSl1p•3/8/23, 8:14 AM

learning rate 0.0002

P1pSl1p•3/8/23, 8:15 AM

but there were a lot of custom settings he walked me through.

P1pSl1p•3/8/23, 8:15 AM

I'll try to compile a .txt file

P1pSl1p•3/8/23, 8:15 AM

just testing the Lora now tho

P1pSl1p•3/8/23, 8:16 AM

checking the flexibility

HolmesCage•3/8/23, 8:32 AM

im trying to generate a safetensors file from my lora but it always saves it like a model and it always has 5gb size what im doing wrong can anyone help me?

vjavad1443•3/8/23, 8:33 AM

@Dr. Furkan Gözükara Hi Dr, I'm truly appreciate for your time and effort. I've learned a lot from your channel & going to implement your tutorial. Just I had a question

Do you think it's better to work on dreambooth by using xformer 0.0.17 nightly or without it ?

vjavad1443•3/8/23, 8:35 AM

Just forgot to say on SD 1.5 I mean

vjavad1443•3/8/23, 8:38 AM

my second question is I wanted to know have you tried the below format for learning rate field instead of 0.00001 ?

0.05:10,0.02:20,0.01:60,0.005:200,0.002:500,0.001:3000,0.0005

I heard it gets better result but wanted to know your opinion as well