Software Engineering Courses (SECourses)•3y ago

@Dr. Furkan Gözükara Hi Dr, I'm truly appreciate for your time and effort. I've learned a lot from y

@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 ?

vjavad1443OP•3/8/23, 8:35 AM

Just forgot to say on SD 1.5 I mean

vjavad1443OP•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

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

P1pSl1p•3/8/23, 8:45 AM

I'll put together a .txt

P1pSl1p•3/8/23, 8:45 AM

just playing around with the lora at the moment

P1pSl1p•3/8/23, 8:46 AM

I've never used a lora before so I'm new to the concept of using one to get accurate results

P1pSl1p•3/8/23, 8:48 AM

P1pSl1p•3/8/23, 8:49 AM

pretty insane if you ask me

gatsby•3/8/23, 8:51 AM

Hey, I want to add the pruned v1.5 and was wondering how I do this? Do I copy the hugging face url and do wget in stable-diffusion-webui / models?

P1pSl1p•3/8/23, 8:51 AM

add it to the models folder in your stable diffusion main

P1pSl1p•3/8/23, 8:52 AM

download it from huggingface

P1pSl1p•3/8/23, 8:52 AM

then move it into your models folder

PP1pSl1p then move it into your models folder

gatsby•3/8/23, 8:53 AM

Thanks, will try that

P1pSl1p•3/8/23, 8:53 AM

P1pSl1p•3/8/23, 8:53 AM

inside models you will see Stable-diffusion

P1pSl1p•3/8/23, 8:53 AM

put it in there

PP1pSl1p I'll put together a .txt

Aurora 神名河•3/8/23, 9:11 AM

Okay thank you!

HolmesCage•3/8/23, 9:15 AM

cant anyone help me :(

AAurora 神名河 Okay thank you!

P1pSl1p•3/8/23, 9:18 AM

realistic vision v1.3

Blip caption photos

batch size 2

LoRA type - Standard

training parameters:

Train Batch size - 2
Epoch - 1
Save every N epoch - 1
caption Extension - .txt

mixed precision - save precision - bf16
number of CPU threads per core - 8
Seed - 1234
cache latent - Check

learning rate - 0.0002
LR Scheduler - constant
LR warmup - 0
Optimizer - AdamW8bit

Text Encoder Learning Rate - 0.00005
Unet learning rate - 0.00008
Network Rank - 128
Network Alpha - 64

Max Rexolution - 768,768
Stop text encoder training - 0
Enable buckets - Check

Advance Settings - Default
Use xformers - Check

P1pSl1p•3/8/23, 9:18 AM

these were the training settings I used

Trimac•3/8/23, 9:22 AM

Can i see ur original photo?

P1pSl1p•3/8/23, 9:23 AM

which original photo? my dataset?

Trimac•3/8/23, 9:23 AM

Yes

Trimac•3/8/23, 9:24 AM

Just few or maybe 2

Trimac•3/8/23, 9:24 AM

im curios

P1pSl1p•3/8/23, 9:24 AM

P1pSl1p•3/8/23, 9:25 AM

and with captions

P1pSl1p•3/8/23, 9:25 AM

P1pSl1p•3/8/23, 9:25 AM

I plan to build a better dataset with more photos and variety

gatsby•3/8/23, 9:27 AM

Does anyone know if when generating class images I can use more than one word? For example instead of “painting” is it okay to use “fine art painting of a woman”

Ggatsby Does anyone know if when generating class images I can use more than one word? F...

P1pSl1p•3/8/23, 9:28 AM

when I tried to train on dreambooth originally with my dataset I had to be more specific in my class image generation because "photo of a woman" didn't give good results for me

PP1pSl1p when I tried to train on dreambooth originally with my dataset I had to be more ...

gatsby•3/8/23, 9:29 AM

Okay, so what did you have to use instead for better results?

P1pSl1p•3/8/23, 9:29 AM

since my wife is thai. I used "photo of a thai woman"

PP1pSl1p Click to see attachment

Trimac•3/8/23, 9:29 AM

Looks like lora copied perfectly your data set but not doing creativity. Well if that what you want then good for you

PP1pSl1p since my wife is thai. I used "photo of a thai woman"

gatsby•3/8/23, 9:30 AM

Ah okay. And then in your instance prompt did you use something like “ohwx photo of a Thai woman”?

TTrimac Looks like lora copied perfectly your data set but not doing creativity. Well i...

P1pSl1p•3/8/23, 9:30 AM

it did a great job for a start. now I need to add more images to the dataset and try again. I'm looking for something very flexible.

P1pSl1p•3/8/23, 9:31 AM

P1pSl1p•3/8/23, 9:31 AM

still good tho

Ggatsby Ah okay. And then in your instance prompt did you use something like “ohwx photo...

P1pSl1p•3/8/23, 9:32 AM

just photo of ohwx woman

PP1pSl1p just photo of ohwx woman

gatsby•3/8/23, 9:32 AM

Thanks so much for your help. Will test it out now

PP1pSl1p Click to see attachment

Trimac•3/8/23, 9:35 AM

Thats good. Can you please try more like a futuristic look prompts? it feels like i wanna try to go back in khoya lora. Im working wd my dreamboot now, wont give me precise face

TTrimac Thats good. Can you please try more like a futuristic look prompts? it feels lik...

P1pSl1p•3/8/23, 9:36 AM

I'll keep at it

P1pSl1p•3/8/23, 9:36 AM

try thos settings I posted above

P1pSl1p•3/8/23, 9:36 AM

training should take aroun 1hr 45

Trimac•3/8/23, 9:37 AM

My god this is so challenging its been months wd training settings but only the old revession dreambooth works once in my life lol! and able to create one model of my face perfectly and thats it Sd broked after that since updates

P1pSl1p•3/8/23, 9:37 AM

the tech is moving so fast

PP1pSl1p try thos settings I posted above

Trimac•3/8/23, 9:38 AM

will do thanks for the share

Aurora 神名河•3/8/23, 9:43 AM

thank you so much for sharing

Aurora 神名河•3/8/23, 9:43 AM

Now LoRA is great again!

P1pSl1p•3/8/23, 9:45 AM

all thanks to @Steel

@Dr. Furkan Gözükara Hi Dr, I'm truly appreciate for your time and effort. I've learned a lot from y

Similar Threads