Software Engineering Courses (SECourses)•13mo ago

As @Dr. Furkan Gözükara himself mentioned, `SUPIR` is way, way, more faithful to the original textur

As @Dr. Furkan Gözükara himself mentioned,

SUPIR

SUPIR

is way, way, more faithful to the original texture than any model that came before it. A

LoRA

LoRA

just rebuilds what it thinks the texture looks like.

Drocsid Imperator MaximusOP•12/11/24, 12:41 PM

That's why I originally went through the trouble. However, I've since had to delete it 3/4 of the way through due to low disk space, heh.

Drocsid Imperator MaximusOP•12/11/24, 12:41 PM

There might be an alternative, but I don't know of one yet.

Drocsid Imperator MaximusOP•12/11/24, 12:42 PM

Not that

JoyCaption

JoyCaption

works either... (after another few GBs too)

fadounav•12/11/24, 1:09 PM

I can't find segmentation sort order @Dr. Furkan Gözükara

Capture_decran_2024-12-11_a_2.08.09_PM.png

DDrocsid Imperator Maximus As @Dr. Furkan Gözükara himself mentioned, `SUPIR` is way, way, more faithful to...

Kallamamran•12/11/24, 1:19 PM

SUPIR is just upscaling. It's not magic

Don't get me wrong. It's great... It just takes a shitload of resources though and it's not fast. SUPIR is like every other "upscaler" nowadays only a "upscaler with an upscaler" together with latent diffusion of the input image. Basically like img2img. "Good", sure, "that good"... Well... I don't think so. It still doesn't "know" anything. It's just guessing, so there's no guarantees the end result will be fantastic. If I can get fantastic results with SUPIR, using ~80GB of drive space and a very long time. I'd rather get fantastic results with ComfyUI TT or SDU in a fraction of the time using models I already have installed on my drive

KKallamamran SUPIR is just upscaling. It's not magic 😊 Don't get me wrong. It's great... It ...

Drocsid Imperator MaximusOP•12/11/24, 1:20 PM

So what's your recommendation then for 8GB VRAM lol?

Kallamamran•12/11/24, 1:20 PM

8GB VRAM and SUPIR doesn't marry that good either huh?

Kallamamran•12/11/24, 1:21 PM

That's quite limiting

Kallamamran•12/11/24, 1:21 PM

Hell I'm struggling with 16GB VRAM on my laptop right now and usually I find 24GB to be on the low end as well

KKallamamran SUPIR is just upscaling. It's not magic 😊 Don't get me wrong. It's great... It ...

fadounav•12/11/24, 1:26 PM

very interesting thank you for sharing your experience

KKallamamran 8GB VRAM and SUPIR doesn't marry that good either huh?

Drocsid Imperator MaximusOP•12/11/24, 1:28 PM

No, I mean another model entirely then with somewhat similar results, heh.

Drocsid Imperator MaximusOP•12/11/24, 1:29 PM

I just wasted 60+GB on a non-working

SUPIR

SUPIR

and

JoyCaption

JoyCaption

FFurkan Gözükara SECourses sorry for delay here

Drocsid Imperator MaximusOP•12/11/24, 1:52 PM

Thanks. Woah!

Drocsid Imperator MaximusOP•12/11/24, 1:52 PM

AuraSR

AuraSR

any better?

Drocsid Imperator MaximusOP•12/11/24, 1:52 PM

It appears to be faster... But I don't know how much space that one takes...

yassinemenace•12/11/24, 1:58 PM

@Dr. Furkan Gözükara When decreasing ddim and alpha should I increase learning rate 5e-5 for 128 what for example 64 ?

fadounav•12/11/24, 3:41 PM

@Dr. Furkan Gözükara I have 2 observations, I have the impression that the generation is long if I compare it to the one on openart in turbo mode. I took 2 A6000, btw, You said in your video that the comparison between lora and finetuning is like night and day. I wanted to know at what level, because if it's rendering quality, I think it's pretty similar. the first image is made with openart and the second with finetuning , it's even worse in finetuning because I've lost the comic illustration aspect.

fadounav•12/11/24, 3:56 PM

1521-illustration_style_of_an_fnfn_man_short-modelTest1-1549214662.jpg

fadounav•12/11/24, 5:17 PM

for info I was able to recover the comic illustration style by adding negative prompts, photorealistic, hyper-realistic, realistic details, etc.

fadounav•12/11/24, 5:25 PM

I think there's only one GPU working, whether it's for one image or 10 in a generation. is that normal? @Dr. Furkan Gözükara

Capture_decran_2024-12-11_a_6.24.26_PM.png

Drocsid Imperator MaximusOP•12/11/24, 5:54 PM

@Dr. Furkan Gözükara Any reliable way to convert

.pt

.pt

.pth

.pth

.bin

.bin

into

.safetensors

.safetensors

Drocsid Imperator MaximusOP•12/11/24, 5:57 PM

I have been testing this, which didn't work correctly, sadly.

GitHub

GitHub - MackinationsAi/convert-2-safetensors: A simple repo for si...

A simple repo for single or batch converting .pt, .pth & .bin models into .safetensors models. - MackinationsAi/convert-2-safetensors

Drocsid Imperator MaximusOP•12/11/24, 6:01 PM

This one now works for some models, but still not all.

GitHub

GitHub - jtabox/safetensors-converter: Converts pth/pt files to saf...

Converts pth/pt files to safetensors. Contribute to jtabox/safetensors-converter development by creating an account on GitHub.

Drocsid Imperator MaximusOP•12/11/24, 6:48 PM

Anybody familiar with this error? Happens with both

.pth

.pth

and

.safetensors

.safetensors

files.

Drocsid Imperator MaximusOP•12/11/24, 6:56 PM

@Dr. Furkan Gözükara Any idea on how to resolve this error?

Drocsid Imperator MaximusOP•12/11/24, 6:58 PM

FYI, useful repo listing all SOTA upscalers sorted by year of release.

GitHub

GitHub - ChaofWang/Awesome-Super-Resolution: Collect super-resoluti...

Collect super-resolution related papers, data, repositories - ChaofWang/Awesome-Super-Resolution

Ffadounav I can't find segmentation sort order @Dr. Furkan Gözükara

Furkan Gözükara SECourses•12/11/24, 8:14 PM

it is never added

Furkan Gözükara SECourses•12/11/24, 8:14 PM

please leave a reply here

Furkan Gözükara SECourses•12/11/24, 8:14 PM

https://github.com/mcmonkeyprojects/SwarmUI/pull/289

GitHub

Add Segmentation Ordering for Yolo segmentation by aimerib · Pull R...

Allow yolo segmentation to sort by left-right, top-bottom, largest-smallest, and reverse that order.
Params added to Regional Segmentation Group

Yyassinemenace @Dr. Furkan Gözükara When decreasing ddim and alpha should I increase learning r...

Furkan Gözükara SECourses•12/11/24, 8:15 PM

dont change alpha and learning rate

Ffadounav @Dr. Furkan Gözükara I have 2 observations, I have the impression that the gener...

Furkan Gözükara SECourses•12/11/24, 8:16 PM

what do you mean by openart?

Ffadounav I think there's only one GPU working, whether it's for one image or 10 in a gene...

Furkan Gözükara SECourses•12/11/24, 8:17 PM

you need to manually add backend

Furkan Gözükara SECourses•12/11/24, 8:17 PM

shown in many tutorials by now

DDrocsid Imperator Maximus @Dr. Furkan Gözükara Any reliable way to convert `.pt` `.pth` `.bin` into `.safe...

Furkan Gözükara SECourses•12/11/24, 8:17 PM

not that i know

DDrocsid Imperator Maximus I just wasted 60+GB on a non-working `SUPIR` and `JoyCaption`

Kallamamran•12/11/24, 8:17 PM

Oh it's working. You must have done something wrong installing

DDrocsid Imperator Maximus @Dr. Furkan Gözükara Any idea on how to resolve this error?

Furkan Gözükara SECourses•12/11/24, 8:17 PM

you should ask the source - i dont know atm

fadounav•12/11/24, 8:20 PM

@Dr. Furkan Gözükara openart is a platform on which I used to create my images before they changed their prices. I used to make fairly high-quality lora flux, after a few image generations on Swarmui I find the quality very good, little waste, but it's a bit slow, I'm going to test with an L40 to go faster what do you think? would it be better with 2A6000?

Ffadounav @Dr. Furkan Gözükara openart is a platform on which I used to create my images ...

Furkan Gözükara SECourses•12/11/24, 8:21 PM

for lora

Furkan Gözükara SECourses•12/11/24, 8:21 PM

you can train 2x gpu

Furkan Gözükara SECourses•12/11/24, 8:21 PM

so yes almost 2x spee

Furkan Gözükara SECourses•12/11/24, 8:22 PM

for fine tuning you cant you need 80 gb gpus

DDrocsid Imperator Maximus No, I mean another model entirely then with somewhat similar results, heh.

Kallamamran•12/11/24, 8:22 PM

Maybe you could get away with a quantized model of Flux, but I really coudn't say. I've tried Q5 and that gives rather good results. I'm not sure if it can be run with 8-steps lora though, so maybe there is no time savings either with that.

FFurkan Gözükara SECourses you can train 2x gpu

Kallamamran•12/11/24, 8:23 PM

https://tenor.com/view/powerful-iron-mask-electricity-shock-zap-gif-18641022

Tenor