Software Engineering Courses (SECourses)•2y ago

even with 1024x1024 variety of 23 images and the face is still trash with adetailer on

ch1253•4/2/24, 1:36 PM

ERROR Invalid user config / ]8;id=184327;file:///kaggle/working/kohya_ss/sd-scripts/library/config_util.py\config_util.py]8;;:]8;id=329280;file:///kaggle/working/kohya_ss/sd-scripts/library/config_util.py#351\351]8;;\

ch1253•4/2/24, 1:36 PM

Can you please suggest?

virginia•4/2/24, 4:08 PM

@Dr. Furkan Gözükara . seeming ONETRAINER is better than koyass , so for wierd shape of products, can ONETRAINDER solve problems, I mean to create images as much as possible like the exact real product

virginia•4/2/24, 4:09 PM

i found many videos saying clothese cannot be trained , because clothes have too many details, and images created cannot show exact real one

Alexm•4/2/24, 4:53 PM

You can most definetly do that, in onetrainer and kohya

Alexm•4/2/24, 4:53 PM

the thing is, getting a dataset good enough to do that

Alexm•4/2/24, 4:53 PM

you need a lot of images showcasing the clothing detail in a lot of different backgrounds, poses, different people and so on

kareem9000•4/2/24, 5:18 PM

@Dr. Furkan Gözükara If a dataset isnt ideal (images a little further, a little pixelated,etc). What would you reccomend to do to it to "clean' it before training?

kareem9000•4/2/24, 5:18 PM

e.g using ur autocropper script is one thing right, what else is something that can be done?

virginia•4/2/24, 5:25 PM

@Alexm really? for example the clothes has several pockets and shows a pattern of several birds here and there,

so like this, such kind of clothes is in too much details

AAIGambino even with 1024x1024 variety of 23 images and the face is still trash with adetai...

Furkan Gözükara SECourses•4/2/24, 5:51 PM

you must be doing something wrong

Furkan Gözükara SECourses•4/2/24, 5:51 PM

i did so many trainings recnetly all worked perfect

Furkan Gözükara SECourses•4/2/24, 5:51 PM

gonna train kohya a new one today

Cch1253 ERROR Invalid user config / ]8;id=184327;file:///kaggle/working/k...

Furkan Gözükara SECourses•4/2/24, 5:52 PM

testing

Vvirginia @Dr. Furkan Gözükara . seeming ONETRAINER is better than koyass , so for wierd s...

Furkan Gözükara SECourses•4/2/24, 5:52 PM

you need to test

Kkareem9000 @Dr. Furkan Gözükara If a dataset isnt ideal (images a little further, a little ...

Furkan Gözükara SECourses•4/2/24, 5:52 PM

crop can fix further ones

Furkan Gözükara SECourses•4/2/24, 5:52 PM

but if quality is low you need to get better ones

Furkan Gözükara SECourses•4/2/24, 5:53 PM

this is how supir download should be completed but dont worry i am uploading as a single file

FFurkan Gözükara SECourses crop can fix further ones

kareem9000•4/2/24, 6:00 PM

gotcha thank you

, is there an optimal number of images? I've seen anywhere from 5 to 15 . i seen some sources saying 12 images, some say 5, etc. (for SDXL training dreambooth)

Kkareem9000 gotcha thank you 🙏 , is there an optimal number of images? I've seen anywhere f...

Furkan Gözükara SECourses•4/2/24, 6:02 PM

more better actually as long as they are higher quality and also diverse

virginia•4/2/24, 6:03 PM

@kareem9000 i think 15-20, 5 is too few

Furkan Gözükara SECourses•4/2/24, 6:03 PM

uploading all supir models as a single file

FFurkan Gözükara SECourses more better actually as long as they are higher quality and also diverse

kareem9000•4/2/24, 6:08 PM

would u say that (generally speaking) if the image dataset is of high quality => its better to use more images when training
but
if the image dataset is not of high quality => better to use less images and (lower resolution training)

Kkareem9000 would u say that (generally speaking) if the image dataset is of high quality =>...

Furkan Gözükara SECourses•4/2/24, 6:10 PM

yes

Furkan Gözükara SECourses•4/2/24, 6:10 PM

never use low resolution

Furkan Gözükara SECourses•4/2/24, 6:10 PM

use at least 1 mega pixels for sdxl

kareem9000•4/2/24, 6:18 PM

yes but at the same time

kareem9000•4/2/24, 6:18 PM

@Dr. Furkan Gözükara in the case where the images are low res (and cant change them) as training date

kareem9000•4/2/24, 6:18 PM

data*

kareem9000•4/2/24, 6:18 PM

then i would be better off training SDXL at lower resolution?

Kkareem9000 then i would be better off training SDXL at lower resolution?

Furkan Gözükara SECourses•4/2/24, 6:21 PM

i think that wouldnt be good

Furkan Gözükara SECourses•4/2/24, 6:21 PM

i would compare supir low text cfg upscaled

Furkan Gözükara SECourses•4/2/24, 6:21 PM

vs raw training

Cch1253 ERROR Invalid user config / ]8;id=184327;file:///kaggle/working/k...

Furkan Gözükara SECourses•4/2/24, 6:25 PM

working

Furkan Gözükara SECourses•4/2/24, 6:25 PM

Furkan Gözükara SECourses•4/2/24, 6:27 PM

Joshcandle•4/2/24, 6:52 PM

@Dr. Furkan Gözükara You said some things will be fixxed in new supir one click installer for runpod right? Do you know when you will release it or do you edit the file in the patreon?

JJoshcandle @Dr. Furkan Gözükara You said some things will be fixxed in new supir one click ...

Furkan Gözükara SECourses•4/2/24, 7:45 PM

runpod is working nice

Furkan Gözükara SECourses•4/2/24, 7:45 PM

if you mean runpod proxy i dont know if fixed

Furkan Gözükara SECourses•4/2/24, 7:45 PM

use gradio live share

Joshcandle•4/2/24, 7:45 PM

No sorry, had some errors last time dont know if you remember

Joshcandle•4/2/24, 7:45 PM

but they seem to be fixed somehow

Furkan Gözükara SECourses•4/2/24, 8:28 PM

Kohya is a little bit faster than OneTrainer when training Stable Diffusion XL (SDXL) with no VRAM optimizations and with half precision BF16

Furkan Gözükara SECourses•4/2/24, 8:28 PM

rafraf•4/2/24, 8:29 PM

@Dr. Furkan Gözükara do you know kohya update « Fix issue with vae path validation » and « Add new setup menu option to install Triton 2.1.0 for Windows » does for training optimisation?

Rrafraf @Dr. Furkan Gözükara do you know kohya update « Fix issue with vae path validati...

Furkan Gözükara SECourses•4/2/24, 8:29 PM

path is working if you mean vae he said fixed

Furkan Gözükara SECourses•4/2/24, 8:29 PM

triton i didnt test

dave5555•4/2/24, 8:39 PM

can anyone reccomend ways to get sam detectors in comfy to recognize non human objects?

dave5555•4/2/24, 8:40 PM

there is a paper about a zero shot sam detection but cant find anything ported to comfy , being able to load an image to guide the sam detection for masking would be awesome

FFurkan Gözükara SECourses path is working if you mean vae he said fixed

rafraf•4/2/24, 8:59 PM

Il just not sure how vae is used in training. Never heard anyone talk about it in training

even with 1024x1024 variety of 23 images and the face is still trash with adetailer on

Similar Threads

Similar Threads

Similar Threads