Software Engineering Courses (SECourses)•14mo ago

What's a good way to test out different checkpoints from a STYLE-BASED finetune? Same seed? poses?

What's a good way to test out different checkpoints from a STYLE-BASED finetune? Same seed? poses?
all of my finetunes look too good.

WWktra What's a good way to test out different checkpoints from a STYLE-BASED finetune?...

Furkan Gözükara SECourses•11/16/24, 11:13 PM

grid

Furkan Gözükara SECourses•11/16/24, 11:13 PM

different prompts

Furkan Gözükara SECourses•11/16/24, 11:13 PM

same seed

WWktra What's a good way to test out different checkpoints from a STYLE-BASED finetune?...

Furkan Gözükara SECourses•11/16/24, 11:13 PM

ye usually all looks good

Furkan Gözükara SECourses•11/16/24, 11:13 PM

with flux

WktraOP•11/16/24, 11:13 PM

different prompts using the same seed?????

WktraOP•11/16/24, 11:13 PM

(I mean SDXL)

WWktra different prompts using the same seed?????

Furkan Gözükara SECourses•11/16/24, 11:14 PM

yes

WWktra (I mean SDXL)

Furkan Gözükara SECourses•11/16/24, 11:14 PM

nice sucess then

WktraOP•11/16/24, 11:15 PM

I must thank you. i used your tutorial to finetune SDXL using a very specific style. (400 training images)
but between epochs 75 - 150, they all look good

WWktra I must thank you. i used your tutorial to finetune SDXL using a very specific st...

Furkan Gözükara SECourses•11/16/24, 11:26 PM

awesome

WktraOP•11/17/24, 6:43 AM

@Furkan Gözükara SECourses
I have trained a finetuned SDXL checkpoint to learn a specific style. The results are excellent.

Now, I would like to finetune that same stylized checkpoint to learn a specific character. Then I will extract it to a LORA.
The character is a female demon. But the training images (40 of them) are in various styles such as 3D, anime style, cartoon style.

If I use the caption "ohwx 1girl" or "ohwx woman", I am afraid that character will OVERWRITE the style of the finetune that I want to keep.
How do I caption the character so that it does not overwrite the style?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/17/24, 9:47 AM

Can't you just avoid the 1girl and woman, and just caption "ohwxgirl", since 1girl and woman are widely recognized tags in any model.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/17/24, 9:48 AM

I did a character finetune, and noticed that almost all training images had ponytail hairstyle. I manage to produce some images without the ponytail, like prompting "messy hair", and even "no ponytail", but the latter didnt really work, haha

WWktra <@205854764540362752> I have trained a finetuned SDXL checkpoint to learn a spe...

Furkan Gözükara SECourses•11/17/24, 11:04 AM

maybe ohwx character

Furkan Gözükara SECourses•11/17/24, 11:05 AM

i dont know you need to test

cyberbol•11/17/24, 11:26 AM

Dr. Furkan. What config you can recommend for me to do fastest way model of face. Not must be a perfect, Priority is time to train on single rtx 4090

cyberbol•11/17/24, 11:26 AM

lora probably as dreambox longer

cyberbol•11/17/24, 11:40 AM

which one i should choose >

Ccyberbol Dr. Furkan. What config you can recommend for me to do fastest way model of face...

Furkan Gözükara SECourses•11/17/24, 11:59 AM

you want lora right?

Ccyberbol which one i should choose >

Furkan Gözükara SECourses•11/17/24, 11:59 AM

it should be faster than 4 second / it

Furkan Gözükara SECourses•11/17/24, 11:59 AM

for you

Furkan Gözükara SECourses•11/17/24, 11:59 AM

like 3 second maybe 2.5

cyberbol•11/17/24, 1:49 PM

I did that. Shared video memory should be used ? Its look like that:

cyberbol•11/17/24, 1:50 PM

281.40s/it

cyberbol•11/17/24, 1:52 PM

cyberbol•11/17/24, 1:52 PM

I think something is wrong

cyberbol•11/17/24, 2:26 PM

I reinstalled everything again. Now I think its ok

cyberbol•11/17/24, 2:26 PM

Ccyberbol Click to see attachment

Furkan Gözükara SECourses•11/17/24, 2:28 PM

now this is a good speed

cyberbol•11/17/24, 2:35 PM

Its possible train on rtx 4090 and 64gb ram using dreambox ? or no way as too slow ?

Ccyberbol Its possible train on rtx 4090 and 64gb ram using dreambox ? or no way as too sl...

Furkan Gözükara SECourses•11/17/24, 2:51 PM

dreambox?

cyberbol•11/17/24, 3:07 PM

DreamBooth

cyberbol•11/17/24, 3:07 PM

Lok

cyberbol•11/17/24, 3:07 PM

Ccyberbol DreamBooth

Furkan Gözükara SECourses•11/17/24, 3:21 PM

yes of course you can train even faster than these speeds

Furkan Gözükara SECourses•11/17/24, 3:21 PM

wardensc2•11/17/24, 3:22 PM

Are there anything new for DreamBooth training @Furkan Gözükara SECourses or only for lora training

Wwardensc2 Are there anything new for DreamBooth training <@205854764540362752> or only for...

Furkan Gözükara SECourses•11/17/24, 3:22 PM

new for both check speeds

wardensc2•11/17/24, 3:22 PM

ah ok

wardensc2•11/17/24, 3:23 PM

but base on your config if I want to get the best quality lora with best speed it is faster to train dreambooth and extract

wardensc2•11/17/24, 3:24 PM

and training time still better than train lora

wardensc2•11/17/24, 3:24 PM

only down side is disk space

cyberbol•11/17/24, 3:37 PM

So at the moment in my situation Lora training is about 3.58s/it, but if I use DreamBooth instead TIER 1 it will be about 7.3s/it??? as I can see on above picture. So double the time. It's worth you think?. Regarding quality?.

Ccyberbol So at the moment in my situation Lora training is about 3.58s/it, but if I use D...

wardensc2•11/17/24, 3:43 PM

tier 1 in dreambooth is different than tier 1 in lora. In lora the rank you get is 128, but in dreambooth I get the maximum rank is 640

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/17/24, 4:18 PM

With 60 training pics for character, is 200 epochs and save every 25 a good setting? I don't think you changed anything about learning rate and other parameters when I watched your tutorial video @Furkan Gözükara SECourses , just use config batch 1? I'm gonna use a80 or h100 later tonight . Want to complete it as fast as possible, with highest possible quality.

cyberbol•11/17/24, 4:29 PM

Oh my God. My computer froze at 48% of the Lora training. Can I somehow resume the training? Or do I have to start over?

cyberbol•11/17/24, 6:30 PM

Now i know we need MARK some option in Koya to be able later resume training

cyberbol•11/17/24, 6:30 PM

Under PArameters - Advence

What's a good way to test out different checkpoints from a STYLE-BASED finetune? Same seed? poses?

Similar Threads