Software Engineering Courses (SECourses)•3y ago

biggest thing when doing captions is to just use the instance prompt, if you use the class prompt, i

ZZono50 class images per instance i only do like 5, if i do too many, it never looks lik...

rafraf•4/30/23, 8:11 PM

interesting. i am also going for photorealism, I've had some pretty interesting results prompting (without my training models) with Dreamlike Photorealism, but not sure, maybe too biased for baking with training... but i will maybe try realisticvision20

Zono50OP•4/30/23, 8:11 PM

and i do 5 sample instance imag4es

Zono50OP•4/30/23, 8:11 PM

yeah same, i'm not big into styles and concepts, i just like doing pretty women mostly

Zono50OP•4/30/23, 8:12 PM

but since they've been getting on my nerves, decided to work on myself today

ZZono50 yeah same, i'm not big into styles and concepts, i just like doing pretty women ...

rafraf•4/30/23, 8:13 PM

i work in fashion, and i was getting results from prompting with a very simple prompt, like this, which is the level of style and photorealism i want.

Zono50OP•4/30/23, 8:13 PM

nice!

ZZono50 and i do 5 sample instance imag4es

rafraf•4/30/23, 8:14 PM

very impressed. I will take note of your settings. will give it a go as well

Zono50OP•4/30/23, 8:17 PM

Anything I can do to help. Lora is a bit of a pain to learn and took me a bit

rafraf•4/30/23, 8:19 PM

so i am training a fashion model i took in studio, and try to render here anywhere, ideally with better resemblance and more expressivity. Haven't used dreambooth for subject and class images yet. this is just from concept LoRA. I need to get up and running. You guys are very inspiring @Zono50 @Furkan Gözükara SECourses

Zono50OP•4/30/23, 8:20 PM

controlnet is great for expressions and poses, doc has tutorials on that

rafraf•4/30/23, 8:20 PM

and this was a LoRA with 5 different outfits separated by subfolders no bleed. Anyway, I am taking you guys' lead to better my training

ZZono50 controlnet is great for expressions and poses, doc has tutorials on that

rafraf•4/30/23, 8:21 PM

yes, that's in the roadmap. I am testing along the lines of Doc and your suggestion with Kohya but I've found StableTuner very cool too, to get live sample updates on my phone's Telegram. Sadly it doesnt have LoRA, so have to do Dreambooth or Finetuninng and extrac LoRA in Kohya

rafraf•4/30/23, 8:21 PM

Anyway, thanks

kashy•4/30/23, 8:24 PM

@Furkan Gözükara SECourses Sorry for the trouble....But i'm banging my head. New to this.

i disabled the previous warning by setting the HF_HUB_DISABLE_SYMLINKS_WARNING environment variable to true.
.
But Even now the problem persists, I even used epoch as 3 (as i read @rafraf 's comment) , but hard luck.
I'm providing a Syntax above.
Any guidance?

message.txt4.42KB

Kkashy @Furkan Gözükara SECourses Sorry for the trouble....But i'm banging my head. New...

Furkan Gözükara SECourses•4/30/23, 8:26 PM

are you trying to run on cpu?

Furkan Gözükara SECourses•4/30/23, 8:27 PM

for cpu, you have to disable all optimizations and dont use fp16 or bf16 or xformers or attention etc

FFurkan Gözükara SECourses are you trying to run on cpu?

kashy•4/30/23, 8:28 PM

Yeah. i have a Vram of 4096 Mib, quite low

FFurkan Gözükara SECourses for cpu, you have to disable all optimizations and dont use fp16 or bf16 or xfor...

kashy•4/30/23, 8:29 PM

shall i set them as No for Mixed Precision and Float for Save precison then?

ZZono50 biggest thing when doing captions is to just use the instance prompt, if you use...

rafraf•4/30/23, 8:31 PM

biggest thing when doing captions is to just use the instance prompt, if you use the class prompt, it seems to get confused, so i run this command in my caption folder - sed -i 's/a man/jrch/g' *.txt - replace man with class prompt and jrch with instance prompt: does this mean that your LoRA caption text file gets just the rare and class tokens, no more? I've been filling the cap txt with descriptives like hair type, direction of face, look on camera off, moles, etc, very specific. you don't?

FFurkan Gözükara SECourses for cpu, you have to disable all optimizations and dont use fp16 or bf16 or xfor...

kashy•4/30/23, 8:39 PM

@Furkan Gözükara SECourses
I disabled like you said.
.
Getting this new error now.
It starts the steps but nothing happens.
and then exits

message.txt7.9KB

Kkashy shall i set them as No for Mixed Precision and Float for Save precison then?

Furkan Gözükara SECourses•4/30/23, 8:59 PM

yep try that

Kkashy @Furkan Gözükara SECourses I disabled like you said. . Getting this new error ...

Furkan Gözükara SECourses•4/30/23, 9:00 PM

you still have mixed precision

Furkan Gözükara SECourses•4/30/23, 9:00 PM

RuntimeError: "slow_conv2d_cpu" not implemented for 'Half'

Furkan Gözükara SECourses•4/30/23, 9:01 PM

can you check all settings to if anyhwere mixed precision

FFurkan Gözükara SECourses can you check all settings to if anyhwere mixed precision

kashy•4/30/23, 9:09 PM

Ok, I used "no" for mixed precision this time,

Its now saying " Error an illegal memory access was encountered at line 167 in file D:\ai\tool\bitsandbytes\csrc\ops.cu "

Sending you a report

message.txt6.96KB

Kkashy Ok, I used "no" for mixed precision this time, Its now saying " Error an illega...

Furkan Gözükara SECourses•4/30/23, 9:22 PM

probably you have to change this optimize

Furkan Gözükara SECourses•4/30/23, 9:22 PM

use 8-bit AdamW optimizer

Furkan Gözükara SECourses•4/30/23, 10:04 PM

i wonder should i make text to speech to improve audio of my videos

FFurkan Gözükara SECourses use 8-bit AdamW optimizer

kashy•4/30/23, 10:14 PM

not happening...Lol. i even reduced the max res to 256,256.
i'll try this with another technique i guess.
I've seen you made a video on that, using google collab.

Kkashy not happening...Lol. i even reduced the max res to 256,256. i'll try this with...

Furkan Gözükara SECourses•4/30/23, 10:16 PM

you can do dreambooth training on shivam colab

Furkan Gözükara SECourses•4/30/23, 10:16 PM

works great

FFurkan Gözükara SECourses you can do dreambooth training on shivam colab

kashy•4/30/23, 10:18 PM

any tutorial/guide you'd know for the same?

Kkashy any tutorial/guide you'd know for the same?

Furkan Gözükara SECourses•4/30/23, 10:18 PM

yes i have 2 tutorials

Furkan Gözükara SECourses•4/30/23, 10:19 PM

you can train there and use on your computer

Furkan Gözükara SECourses•4/30/23, 10:19 PM

1st : https://youtu.be/mnCY8uM7E50

YouTubeSECourses

Transform Your Selfie into a Stunning AI Avatar with Stable Diffusi...

Our Discord : https://discord.gg/HbqgGaZVmr. How to do free Stable Diffusion DreamBooth training on Google Colab for free. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials, Automatic1111 and Goo...

Furkan Gözükara SECourses•4/30/23, 10:19 PM

2 nd : https://youtu.be/kIyqAdd_i10

YouTubeSECourses

Stable Diffusion Google Colab, Continue, Directory, Transfer, Clone...

Our Discord : https://discord.gg/HbqgGaZVmr. This is the video where you will learn how to use Google Colab for Stable Diffusion. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials, #Automatic1111...

Furkan Gözükara SECourses•4/30/23, 10:20 PM

dreambooth still really powerful when compared to lora

Furkan Gözükara SECourses•4/30/23, 10:20 PM

this is dreambooth - didnt use face restore

Furkan Gözükara SECourses•4/30/23, 10:20 PM

FFurkan Gözükara SECourses yes i have 2 tutorials

kashy•4/30/23, 10:21 PM

Thankyou Sir! It's amazing what u do. Inspired!

Kkashy Thankyou Sir! It's amazing what u do. Inspired!

Furkan Gözükara SECourses•4/30/23, 10:21 PM

you are welcome

Furkan Gözükara SECourses•4/30/23, 10:21 PM

making same query for lora now

FFurkan Gözükara SECourses making same query for lora now

kashy•4/30/23, 10:23 PM

Thanks, It'll be helpful. I just read your comment on someone's query on the lora video, that it's not possible on 4gb Vram, wish i had read earlier. haha, great scratching my head, learned a lot though.

Kkashy Thanks, It'll be helpful. I just read your comment on someone's query on the lor...

Furkan Gözükara SECourses•4/30/23, 10:24 PM

ye 4 gb not possible :/

Isaac Brock•4/30/23, 10:34 PM

what is the significance of the loss value during lora training?

Isaac Brock•4/30/23, 10:34 PM

is a lower value more desirable?

Isaac Brock•4/30/23, 10:34 PM

im finding that i need to redo training 3 or 4x before getting a good result

Isaac Brock•4/30/23, 10:35 PM

like restoring from the previous state, not fresh every time

Isaac Brock•4/30/23, 10:35 PM

so a total of like 15-20k steps

Furkan Gözükara SECourses•4/30/23, 10:45 PM

lower loss is desired but doesnt mean well training aone

biggest thing when doing captions is to just use the instance prompt, if you use the class prompt, i

Similar Threads