Software Engineering Courses (SECourses)•16mo ago

unfortunately it also happens with no upscaling / also with lower resolution like 1024x512 When I ru

unfortunately it also happens with no upscaling / also with lower resolution like 1024x512
When I run the same prompt and seed without the lora there are no vertical lines
Is my Lora "broken"?

FFurkan Gözükara SECourses use our config dont change it dont load into dreamboth tab even a once

Phuong Bui•9/7/24, 9:22 PM

i use a40 what rank should i use ?

LLuciani unfortunately it also happens with no upscaling / also with lower resolution lik...

Furkan Gözükara SECourses•9/7/24, 9:25 PM

then very likely

PPhuong Bui i use a40 what rank should i use ?

Furkan Gözükara SECourses•9/7/24, 9:25 PM

rank 1 best

Furkan Gözükara SECourses•9/7/24, 9:25 PM

but it is too slow if you need speed

Furkan Gözükara SECourses•9/7/24, 9:25 PM

rank 2

Furkan Gözükara SECourses•9/7/24, 9:26 PM

i feel like with consistent dataset no captions worked better than captions

Furkan Gözükara SECourses•9/7/24, 9:26 PM

for style for flux training

Furkan Gözükara SECourses•9/7/24, 9:26 PM

let me share link with you

LucianiOP•9/7/24, 9:27 PM

i did not use any caption files and also no reference images -- just 100 epochs with repeat set to 1 having 132 images to train

LucianiOP•9/7/24, 9:29 PM

however i notice the vertical lines are "lesser" with lower initial resolution like 1024x512 (instead of something like 1400x700)

LucianiOP•9/7/24, 9:29 PM

what generation resolution do you recommend in common?

Furkan Gözükara SECourses•9/7/24, 9:33 PM

https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX/resolve/main/Fixed-Consistent-Training-Dataset-Results-Grid-50700x15500px.jpg

LLuciani however i notice the vertical lines are "lesser" with lower initial resolution l...

Furkan Gözükara SECourses•9/7/24, 9:34 PM

make all train iamges 1024x1024 and generate 1024x1024

Furkan Gözükara SECourses•9/7/24, 9:34 PM

then you can try multi resolution

LLuciani i did not use any caption files and also no reference images -- just 100 epochs ...

Furkan Gözükara SECourses•9/7/24, 10:00 PM

currently people reporting multi resolution with kohya not working very good for flux - i am not sure - but as i said first go with 1024x1024 and once you have some decent results you can try multi res training and compare

LucianiOP•9/7/24, 10:06 PM

or maybe it is because of my pics are being oversized for training? Or is it rather the mixed aspect ratio?

LLuciani or maybe it is because of my pics are being oversized for training? Or is it rat...

Furkan Gözükara SECourses•9/7/24, 10:09 PM

mixed aspect ratio

Furkan Gözükara SECourses•9/7/24, 10:09 PM

you have too many

Furkan Gözükara SECourses•9/7/24, 10:09 PM

and too little number of images

LucianiOP•9/7/24, 10:10 PM

I have 132 images for lora training -- I thought that even would be too much -- as other people I heard of only train with like 20 images? So 132 is commonly ok?

FFurkan Gözükara SECourses rank 1 best

Phuong Bui•9/7/24, 10:13 PM

I used rank 2 and it worked. But the problem is that I used 25 images but I don't know why the num train images * repeats is up to 1025. Am I setting it wrong somewhere? And if so, where should I go to fix it?

LLuciani I have 132 images for lora training -- I thought that even would be too much -- ...

Furkan Gözükara SECourses•9/7/24, 10:15 PM

for flux

Furkan Gözükara SECourses•9/7/24, 10:15 PM

more better

PPhuong Bui I used rank 2 and it worked. But the problem is that I used 25 images but I don'...

Furkan Gözükara SECourses•9/7/24, 10:15 PM

make repeat 1

Furkan Gözükara SECourses•9/7/24, 10:15 PM

if you have another folder in training folder

Furkan Gözükara SECourses•9/7/24, 10:15 PM

it uses all

FFurkan Gözükara SECourses make repeat 1

Phuong Bui•9/7/24, 10:16 PM

thanks you

DEXperience Solutions•9/7/24, 10:17 PM

I just made the payment, and I hope you can help me achieve the goal of generating comic images of my NFTs. I’m starting to watch your video. Best regards.

DDEXperience Solutions I just made the payment, and I hope you can help me achieve the goal of generati...

Furkan Gözükara SECourses•9/7/24, 10:21 PM

welcome

Furkan Gözükara SECourses•9/7/24, 10:21 PM

read here very carefully : https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX

MonsterMMORPG/3D-Cartoon-Style-FLUX · Hugging Face

Furkan Gözükara SECourses•9/7/24, 10:21 PM

style training concluded i written everything there

LucianiOP•9/7/24, 10:26 PM

So to ensure -- it would be ok to have mixed resolution for training images but they should have the same aspect ratio

For example havin mixed 1000x1000, 1200x1200, 1400x1400 and so on but all have the aspect ratio of 1:1
Can I have the some with aspect ratio 2:1 like havin 70 images with 2000x1000 and 70 with 1000x1000?

LucianiOP•9/7/24, 10:27 PM

Background: I do not want always only to crop the face -- because the lora should also "learn" the body proportions

LLuciani So to ensure -- it would be ok to have mixed resolution for training images but ...

Furkan Gözükara SECourses•9/7/24, 10:32 PM

this sounds better than your older case

LucianiOP•9/7/24, 10:33 PM

i´ll try, thank you -- by the way - I also subscribed to your patreon cause I think you do a great work, thank you!

Furkan Gözükara SECourses•9/7/24, 10:56 PM

our style model published with full details : https://civitai.com/models/731347

SECourses 3D Render for FLUX - Full Dataset and Workflow Shared - v...

Full Training Tutorial and Guide and Research For a FLUX Style Hugging Face repo with all checkpoints, prompts and more details > https://huggin...

DEXperience Solutions•9/7/24, 11:00 PM

@Dr. Furkan Gözükara I’ve read in detail what you did with MonsterMMORPG/3D-Cartoon-Style-FLUX, and now I’ll try to train following that guide. I have 800 images of 2608 x 2608; they are all spheres that vary in accessories and country color. My goal, after achieving a decent result that reuses the character in different scenarios, is to deploy it on an API that I will later query through a Telegram bot. One question I have is whether to use Massed Compute or RunPod?

DDEXperience Solutions @Dr. Furkan Gözükara I’ve read in detail what you did with MonsterMMORPG/3D-Cart...

Furkan Gözükara SECourses•9/7/24, 11:04 PM

go massed compute

Furkan Gözükara SECourses•9/7/24, 11:04 PM

way more dependable

Furkan Gözükara SECourses•9/7/24, 11:04 PM

rent 4x A6000 - 1.25 usd per hour

Furkan Gözükara SECourses•9/7/24, 11:04 PM

probably you can get good results with early epochs if they are consistent

Furkan Gözükara SECourses•9/7/24, 11:05 PM

66 images required only 75 epochs

DEXperience Solutions•9/7/24, 11:07 PM

I have 800 different characters, between accessories and country, do you think that training with the 800 images is not the way to go?
I should choose 100 and work with them?
Because per country I have 4-6 versions which share the same flag color, but different accessories.

DEXperience Solutions•9/7/24, 11:09 PM

The least I hope in this experiment is to get to occupy a prompt, that I can pass it for example “Ball on the beach” I hope one of those balls with a background on the beach.

DDEXperience Solutions I have 800 different characters, between accessories and country, do you think t...

Furkan Gözükara SECourses•9/7/24, 11:09 PM

as long as they consistent and not repeating i would say use all

Furkan Gözükara SECourses•9/7/24, 11:09 PM

also in your case captions may work better not sure though

DDEXperience Solutions I have 800 different characters, between accessories and country, do you think t...