Software Engineering Courses (SECourses)•3y ago

Is there a youtube video or something how to use this kind of caption for Lora? @Dr. Furkan Gözükara

Marcelo•10/22/23, 5:30 AM

Sdxl quality is insane, thank you for everything @Dr. Furkan Gözükara . Its 2 pm where I live and I managed to successfully create my very first high quality model. I had to tweak the learning rate at least 4 times, because it was overtraining too fast, but the final result is incredible. Thank u so much for your videos, tutorials and scripts, glad to be a patreon supporter :100_rainbow:

MMarcelo Sdxl quality is insane, thank you for everything @Dr. Furkan Gözükara . Its 2 pm...

Benjamin•10/22/23, 6:50 AM

what learning rate did you end up with?

FFurkan Gözükara SECourses https://twitter.com/GozukaraFurkan/status/1715874999204982896

Benjamin•10/22/23, 6:52 AM

great news

FFurkan Gözükara SECourses updated for sdxl dreambooth on kaggle - free account : https://www.patreon.com/p...

Benjamin•10/22/23, 7:01 AM

is it maybe possible to save more checkpoints if we put them in a temporary folder like we used to do with the downloaded models?

LEGACIES•10/22/23, 7:07 AM

@Dr. Furkan Gözükara is it more accurate to crop my dataset with your auto crop script compared to cropping it with photoshop?

LEGACIES•10/22/23, 7:29 AM

How to enable bucketing?

mikemenders•10/22/23, 8:44 AM

In kohya here:

mikemenders•10/22/23, 10:27 AM

I just trained a six-keyword Lora for SD 1.5 (poses). And it's nice to see that it produces results for all six keywords nicely, and just need a minimal Hires fix (0.3) for face. With AdamW8bit optimizer, with these settings (almost the same as I shared earlier, slightly modified and higher weights).
Now I'm retraining with the same parameters, but I've turned on normalization, which also gave interesting results yesterday, but as I saw, it does match the Lora to the base model. I'll be curious to see what I end up with.

Mykee_-_AdamW_-_SD15_-_Concept_best.json2.79KB

mikemenders•10/22/23, 10:40 AM

@Dr. Furkan Gözükara check this PR

https://github.com/kohya-ss/sd-scripts/pull/893

GitHub

Add EMA by vvern999 · Pull Request #893 · kohya-ss/sd-scripts

Add EMA support. Usually it is used like this:

At the start of training, the model parameters are copied to the EMA parameters.
At each gradient update step, the EMA parameters are updated using t...

BBenjamin what learning rate did you end up with?

Marcelo•10/22/23, 10:57 AM

2e-05

Sandstormleader•10/22/23, 12:31 PM

@Dr. Furkan Gözükara do we have an up to date version of dreambooth and automatic1111 that is working yet or do i still need to train from older version?

Sandstormleader•10/22/23, 12:32 PM

was hoping to try training my first sdxl model today

Mmikemenders @Dr. Furkan Gözükara check this PR 😉 https://github.com/kohya-ss/sd-scripts/pul...

Benjamin•10/22/23, 1:02 PM

maybe interesting for you @mikemenders : https://github.com/bghira/SimpleTuner/releases/tag/v0.7.0

GitHub

Release v0.7.0 - tune SDXL in just 12G VRAM! · bghira/SimpleTuner

DeepSpeed lives!
Now correctly integrated, DeepSpeed allows for training SDXL's full U-net at 8 seconds per iteration on just ~12G of VRAM. See the documentation for more information!
Features
...

sharkdodo•10/22/23, 1:06 PM

a question on lora with various checkpoints: in the previous sd1.5 era, i could train a lora on sd1.5 or anylora checkpoint and use that lora on almost any checkpoints without fail. However, loras trained on & tested ok with sdxl model doesnt seem to work at all with other checkpoints (such as realVisXL) - it generates a lot of artifacts as if overtrained. Does anyone have idea why?

Sandstormleader•10/22/23, 1:31 PM

anyone have a link to the automatic1111 sdxl auto installer for windows?

Abolfazl•10/22/23, 3:57 PM

لطفا به فارسی چت کن

Abolfazl•10/22/23, 3:57 PM

سلام

Abolfazl•10/22/23, 3:58 PM

Hello

GGeorgeClintonGeorgeBush What's the current best SDXL checkpoint for realistic photos of people?

Furkan Gözükara SECourses•10/22/23, 4:35 PM

sdxl 1.0 really good

Furkan Gözükara SECourses•10/22/23, 4:35 PM

havent compared others yet

Pphillip Is there a youtube video or something how to use this kind of caption for Lora? ...

Furkan Gözükara SECourses•10/22/23, 4:35 PM

i dont have yet but you can test each one do a train and compare

Furkan Gözükara SECourses•10/22/23, 4:35 PM

best approach

MMarcelo Sdxl quality is insane, thank you for everything @Dr. Furkan Gözükara . Its 2 pm...

Furkan Gözükara SECourses•10/22/23, 4:36 PM

awesome thank you

DR.Siva•10/22/23, 4:46 PM

ok it seems lobe theme needs internet to properly load

BBenjamin is it maybe possible to save more checkpoints if we put them in a temporary fold...

Furkan Gözükara SECourses•10/22/23, 4:47 PM

yes but kaggle has a problem right now with temp folder

Furkan Gözükara SECourses•10/22/23, 4:47 PM

i got reply from kaggle employee

Furkan Gözükara SECourses•10/22/23, 4:47 PM

it works super slow

LLEGACIES @Dr. Furkan Gözükara is it more accurate to crop my dataset with your auto crop ...

Furkan Gözükara SECourses•10/22/23, 4:47 PM

i think if you crop same way would be same

Furkan Gözükara SECourses•10/22/23, 4:47 PM

but you dont need to waste time

Ssharkdodo a question on lora with various checkpoints: in the previous sd1.5 *era*, i coul...

DR.Siva•10/22/23, 4:48 PM

yes i too felt the same. Lora`s dont play well with models imo. Also i found , setting them to a 0.9 or less than that results in much controlled output.

Mmikemenders @Dr. Furkan Gözükara check this PR 😉 https://github.com/kohya-ss/sd-scripts/pul...

Furkan Gözükara SECourses•10/22/23, 4:48 PM

amazing news

Furkan Gözükara SECourses•10/22/23, 4:48 PM

after this sd 1.5 can be researched throughly

SSandstormleader @Dr. Furkan Gözükara do we have an up to date version of dreambooth and automati...

Furkan Gözükara SECourses•10/22/23, 4:48 PM

older working only

Furkan Gözükara SECourses•10/22/23, 4:48 PM

but after EMA of kohya we won't need it

DR.Siva•10/22/23, 4:49 PM

qwen-flux-wan-such Will try this soon..

BBenjamin maybe interesting for you @mikemenders : https://github.com/bghira/SimpleTuner/r...

Furkan Gözükara SECourses•10/22/23, 4:49 PM

i think instead using kaggle better

Furkan Gözükara SECourses•10/22/23, 4:50 PM

you can train with 2 hours 160 epochs

FFurkan Gözükara SECourses i think if you crop same way would be same

LEGACIES•10/22/23, 4:50 PM

how can i use bucket in dreambooth extension for sd 1.5?

Ssharkdodo a question on lora with various checkpoints: in the previous sd1.5 *era*, i coul...

Furkan Gözükara SECourses•10/22/23, 4:50 PM

because sd 1.5 models are vastly merge

LLEGACIES how can i use bucket in dreambooth extension for sd 1.5?

Furkan Gözükara SECourses•10/22/23, 4:50 PM

i think sd 1.5 dont have bucket option

Furkan Gözükara SECourses•10/22/23, 4:51 PM

or it is doing auto

Furkan Gözükara SECourses•10/22/23, 4:51 PM

cant remember

SSandstormleader anyone have a link to the automatic1111 sdxl auto installer for windows?

Furkan Gözükara SECourses•10/22/23, 4:51 PM

here https://www.patreon.com/posts/86307255

LEGACIES•10/22/23, 4:53 PM

i saw in your video that for auto cropping script that you have sent the dataset into the train tabl to resize it, do i need to do the same with my dataset after cropping it 512px x 768pxt?

LEGACIES•10/22/23, 4:54 PM

LLEGACIES i saw in your video that for auto cropping script that you have sent the dataset...

Furkan Gözükara SECourses•10/22/23, 4:57 PM

auto1111 cropping so bad

Furkan Gözükara SECourses•10/22/23, 4:57 PM

dont use

Furkan Gözükara SECourses•10/22/23, 4:57 PM

use my script

Sandstormleader•10/22/23, 5:35 PM

is kohya something you run locally?

SSandstormleader is kohya something you run locally?

Furkan Gözükara SECourses•10/22/23, 5:42 PM

yes

Is there a youtube video or something how to use this kind of caption for Lora? @Dr. Furkan Gözükara

Similar Threads