Software Engineering Courses (SECourses)•17mo ago

I'm wondering if its even doing anything. Because for example i tried "ohwx womas as snow white" and

I'm wondering if its even doing anything. Because for example i tried "ohwx womas as snow white" and the lora is not even activated, just creates a cartoon image of snow white exactly how snow white is despite lora weight 100% and photo real images on same lora result in very close likeness.

It seems like the lora is only activating on photo real images, and at that point adding ohwx does nothing to change the image,

so the flux loras right now just seems to work by replace all men/women with trained person when its generating photo real image with a man/woman.

Darth Teethius•8/20/24, 8:33 PM

I read on reddit there are parameters realized by kohya that bring lora flux training under 12gb vram. I cant find anything of the like. Anyone seen these alleged parameters?

Cclarkcheekykent I'm wondering if its even doing anything. Because for example i tried "ohwx woma...

Furkan Gözükara SECourses•8/20/24, 8:47 PM

I can try prompt what is it exactly?

DDarth Teethius I read on reddit there are parameters realized by kohya that bring lora flux tra...

Furkan Gözükara SECourses•8/20/24, 8:47 PM

Those param Currently using 18gb instead of 12 I have reported

daroth•8/20/24, 8:50 PM

any of you tried out the civitai onsite trainer for flux?

daroth•8/20/24, 8:54 PM

what the hell could be the problem

FFurkan Gözükara SECourses I can try prompt what is it exactly?

clarkcheekykentOP•8/20/24, 8:58 PM

it was "ohwx woman as snow white" but i guess you can try maybe "[trigger] man as Aladdin" instead since your Lora would be male.
If its a photoreal image, add "cartoon image of [trigger] man as Aladdin"

FFurkan Gözükara SECourses haha 200 epoch high LR completely memorized

Haashim•8/20/24, 9:03 PM

How did you caption the images?

Ddaroth what the hell could be the problem

clarkcheekykentOP•8/20/24, 9:04 PM

did you already click submit? just refresh page on the training page it should start. I had trained one lora already.

daroth•8/20/24, 9:04 PM

I managed to resubmit it with the default training params. not sure how well dim2 will work to be honest

HHaashim How did you caption the images?

Furkan Gözükara SECourses•8/20/24, 9:05 PM

Ohwx man

Cclarkcheekykent it was "ohwx woman as snow white" but i guess you can try maybe "[trigger] man a...

Furkan Gözükara SECourses•8/20/24, 9:05 PM

But this prompt doesn't make sense

daroth•8/20/24, 9:06 PM

just ohwx man? I thought flux likes natural language caps

FFurkan Gözükara SECourses But this prompt doesn't make sense 🙂

clarkcheekykentOP•8/20/24, 9:11 PM

i meant "ohwx man as Aladdin", why does it not make sense?

for example here is "Joe Biden as Aladdin"

Dazzastrous [4090]•8/20/24, 9:11 PM

hahha

Dazzastrous [4090]•8/20/24, 9:11 PM

and wheres the Aladdin version?

clarkcheekykentOP•8/20/24, 9:12 PM

It attempted to at least generate Joe Biden in Aladdin cartoon style with some features lke the colors similar to Aladdin style. What I'm seeing with my Lora is it just creates snow white exactly as she is in cartoons instead of with my persons likeness

Dazzastrous [4090]•8/20/24, 9:14 PM

have you seen five below zero's lora training quite informative too but hes a comfyui

clarkcheekykentOP•8/20/24, 9:14 PM

it works with "Hilary Clinton as Snow White"

Dazzastrous [4090]•8/20/24, 9:14 PM

very good

Dazzastrous [4090]•8/20/24, 9:15 PM

now we get to see what bill sees behind closed doors

Dazzastrous [4090]•8/20/24, 9:15 PM

take a bite yeh enjoy that appleeee

Dazzastrous [4090]•8/20/24, 9:16 PM

still im glad you cropped him out bending over

dxqb•8/20/24, 9:23 PM

what the best UI currently to try out flux?

Cclarkcheekykent it works with "Hilary Clinton as Snow White"

Furkan Gözükara SECourses•8/20/24, 9:29 PM

2127-photo20of20ohwx20man20as20aladdinsegmentfac-flux1-dev-1241851401.png

Furkan Gözükara SECourses•8/20/24, 9:29 PM

here me as aladdin

Ddxqb what the best UI currently to try out flux?

Furkan Gözükara SECourses•8/20/24, 9:29 PM

swarm

FFurkan Gözükara SECourses Click to see attachment

Dazzastrous [4090]•8/20/24, 9:30 PM

hehehe

sexyboi•8/20/24, 9:30 PM

@Dr. Furkan Gözükara what are your current impressions on the necessary training time of flux vs sdxl?

Ssexyboi @Dr. Furkan Gözükara what are your current impressions on the necessary training...

Furkan Gözükara SECourses•8/20/24, 9:30 PM

i am testing same atm same dataset

Furkan Gözükara SECourses•8/20/24, 9:30 PM

still in research

Dazzastrous [4090]•8/20/24, 9:30 PM

you want quality you gotta wait a lil longer

DDazzastrous [4090]you want quality you gotta wait a lil longer

Furkan Gözükara SECourses•8/20/24, 9:31 PM

fine tuning will have better quality

Furkan Gözükara SECourses•8/20/24, 9:31 PM

this is lora

Furkan Gözükara SECourses•8/20/24, 9:31 PM

i will next research fine tuning

Dazzastrous [4090]•8/20/24, 9:32 PM

although Ive seen some train lora on flux with around 13 epochs

Dazzastrous [4090]•8/20/24, 9:32 PM

11 to 13 seems the sweet spot

FFurkan Gözükara SECourses fine tuning will have better quality

Dazzastrous [4090]•8/20/24, 9:33 PM

looks pretty good to me alrteady Doc

FFurkan Gözükara SECourses fine tuning will have better quality

dxqb•8/20/24, 9:34 PM

OneTrainer dev is making the claim that you can have finetuning quality when using Dora-Loras, that they have recently implemented and I think Kohya had for a while

dxqb•8/20/24, 9:34 PM

I've not tested it, just might be something interesting for you

Dazzastrous [4090]•8/20/24, 9:35 PM

Dora Loras hehe just rolls of the tongue

dxqb•8/20/24, 9:35 PM

it's only called DoRA, I just added LoRA to make clear that it's just another type of lora

Dazzastrous [4090]•8/20/24, 9:36 PM

All we need is Sora Dora Loras

DDazzastrous [4090]All we need is Sora Dora Loras

dxqb•8/20/24, 9:37 PM

paper by Liu, Wang, Yin andChen

dxqb•8/20/24, 9:37 PM

https://arxiv.org/abs/2402.09353

arXiv.org

DoRA: Weight-Decomposed Low-Rank Adaptation

Among the widely used parameter-efficient fine-tuning (PEFT) methods, LoRA and its variants have gained considerable popularity because of avoiding additional inference costs. However, there still often exists an accuracy gap between these methods and full fine-tuning (FT). In this work, we first introduce a novel weight decomposition analysis t...