Software Engineering Courses (SECourses)•14mo ago

if you turn down guidence you can stilize really well

FFelipe Machado if you turn down guidence you can stilize really well

Furkan Gözükara SECourses•11/11/24, 2:48 PM

nice

Xcorpion•11/12/24, 12:48 AM

just trained a lora on a recent sdxl model (noobAI V-pred0.5), and it turned out like this. but some loras trained on epsilon versions works with vpred checkpoints.
this is just a speculation, since the checkpoint is not epsilon structure (dont know what is epsilon or vpred meaning) kohya can't train it or smthing?

03523-451806595-masterpiece20best20quality20absurdres20highly20detailed_lora_style1noobai0.png

XXcorpion just trained a lora on a recent sdxl model (noobAI V-pred0.5), and it turned out...

Furkan Gözükara SECourses•11/12/24, 12:52 AM

very likely. try to train on sdxl base first

Furkan Gözükara SECourses•11/12/24, 12:52 AM

and see if your config works

Xcorpion•11/12/24, 6:05 AM

@Furkan Gözükara SECourses a friend asked me to train using his gpu too, so is it possible in kohya or in any other way? I got 4070, he got 4060ti.

Xcorpion•11/12/24, 6:07 AM

and I see these checkpoint models trained on 2x A100 GPU, 4x A100 GPU... so we can combine gpus or smthing?

XXcorpion <@205854764540362752> a friend asked me to train using his gpu too, so is it pos...

Furkan Gözükara SECourses•11/12/24, 11:47 AM

nope you cant combine

Furkan Gözükara SECourses•11/12/24, 11:47 AM

like that

Furkan Gözükara SECourses•11/12/24, 11:47 AM

on linux 2x 3090 may work for lora

Felipe MachadoOP•11/13/24, 1:51 PM

hello again, how do I resume a Lora training in Khoya? I know how to pickup a training I left when training a finetuned checkpoint by selecting the latest saved safetensor in the " Pretrained model name or path" but I dont know hot to resume a previews Lora training , thanks for the help

FFelipe Machado hello again, how do I resume a Lora training in Khoya? I know how to pickup a tr...

Furkan Gözükara SECourses•11/13/24, 1:56 PM

tips-info

wardensc2•11/15/24, 1:03 AM

Hi @Furkan Gözükara SECourses have you ever train hyper flux or turbo flux model

wardensc2•11/15/24, 1:04 AM

I trained model and merge with hyper or turbo 8 step lora. The image result for 8 step for normal gen is ok but the upscale quality is worst than normal

wardensc2•11/15/24, 1:06 AM

I wonder if i train on an already merged hyper or turbo does image will be better or not

Wwardensc2 Hi <@205854764540362752> have you ever train hyper flux or turbo flux model

Furkan Gözükara SECourses•11/15/24, 12:19 PM

nope i havent

yassinemenace•11/15/24, 4:39 PM

@Furkan Gözükara SECourses do you do face inpainting for close up photos ?

yassinemenace•11/15/24, 4:40 PM

Sometimes it messes up with eyes looking direction

yassinemenace•11/15/24, 4:40 PM

What do you think ?

Yyassinemenace <@205854764540362752> do you do face inpainting for close up photos ?

Furkan Gözükara SECourses•11/15/24, 8:01 PM

i skip in some poses

Furkan Gözükara SECourses•11/15/24, 8:01 PM

if it generates too close

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/16/24, 4:44 PM

If my pics are jpg, should they be converted to png? Which tool should I use if I have a jpg file with say, 300 X 300 px to make it a good training image l?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/16/24, 4:50 PM

@Furkan Gözükara SECourses

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭If my pics are jpg, should they be converted to png? Which tool should I use if ...

Furkan Gözükara SECourses•11/16/24, 4:54 PM

300x300 is terrible

Furkan Gözükara SECourses•11/16/24, 4:54 PM

jpg png convert doesnt matter

Furkan Gözükara SECourses•11/16/24, 4:54 PM

but if original image png better

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/16/24, 8:57 PM

ok thank you. @Furkan Gözükara SECourses I am just about to start dreambooth training, and I want the best quality. I heard you make a note that batch 1 48GB (using a6000) actually yields better quality than batch 7, so the only advantage with batch 7 is the speed. Would you recommend batch 1 then if I want the best quality?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/16/24, 9:03 PM

I am using 66 training images. I went for 125 epochs and save every 25 epoch.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/16/24, 9:10 PM

ok trying out batch 7 first as you suggested, if quality is not good enough, I might run batch 1 later. By the way, Why either 15 or 256 images, is that some sort of magic numbers for fine-tuning?

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭ok thank you. <@205854764540362752> I am just about to start dreambooth training...

Furkan Gözükara SECourses•11/16/24, 10:22 PM

yep batch size 1

Furkan Gözükara SECourses•11/16/24, 10:22 PM

i recommend

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭ok trying out batch 7 first as you suggested, if quality is not good enough, I m...

Furkan Gözükara SECourses•11/16/24, 10:23 PM

nope it was just a big comparison

Furkan Gözükara SECourses•11/16/24, 10:23 PM

so that you can know between both works great :d

Furkan Gözükara SECourses•11/16/24, 10:23 PM

256 too many but still working

Furkan Gözükara SECourses•11/16/24, 10:23 PM

with same config

Wktra•11/16/24, 10:58 PM

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

Wktra•11/16/24, 11:13 PM

different prompts using the same seed?????

Wktra•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

Wktra•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

Wktra•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

if you turn down guidence you can stilize really well

Similar Threads

Similar Threads

Similar Threads