Software Engineering Courses (SECourses)•10mo ago

Regarding upscaling in swarmui, what is the 4xrealwebphoto model that you use in tips-info @Dr. Furk

Regarding upscaling in swarmui, what is the 4xrealwebphoto model that you use in tips-info @Dr. Furkan Gözükara ? Or which one is the best to use (for realistic portrait photos)

FFurkan Gözükara SECourses awesome

HRK•2/28/25, 2:37 PM

interestingly enough, training a ~250 image dataset on wan2.1 14b "only" takes 27gb gpu at res 1024 and rank 128, I wonder how far it can be pushed with a 80gb gpu

HRK•2/28/25, 2:38 PM

but to be fair it's quite slow per iteration (~23sec)

HRK•2/28/25, 2:39 PM

need to try videos next but I really want to push that training to the max

HRK•2/28/25, 3:12 PM

BUT the lora is 1gb damn

TTaz Found it on X. How true not sure, but need more investigation.

Furkan Gözükara SECourses•2/28/25, 3:42 PM

this guy doesnt know

Furkan Gözükara SECourses•2/28/25, 3:42 PM

he confused telemetry data

Jjulius Regarding upscaling in swarmui, what is the 4xrealwebphoto model that you use in...

Furkan Gözükara SECourses•2/28/25, 3:54 PM

it is here

Furkan Gözükara SECourses•2/28/25, 3:54 PM

https://openmodeldb.info/

OpenModelDB

OpenModelDB is a community driven database of AI Upscaling models. We aim to provide a better way to find and compare models than existing sources.

Furkan Gözükara SECourses•2/28/25, 3:54 PM

my unified downloader also downloads it

HHRK interestingly enough, training a ~250 image dataset on wan2.1 14b "only" takes 2...

Furkan Gözükara SECourses•2/28/25, 3:54 PM

how good it works?

Furkan Gözükara SECourses•2/28/25, 3:54 PM

i plan make a gui for this

Furkan Gözükara SECourses•2/28/25, 3:55 PM

by the way sd 3.5 large training getting improved. now we gonna have trained clip l and clip g into embedded file

FFurkan Gözükara SECourses how good it works?

HRK•2/28/25, 3:56 PM

I'm still training, it's very slow but it works well on a A6000, I've not trained videos yet but I will work on it this week end, but it's promising tbh, the lora file is 1gb tho, I'm waiting for a few epoch and will test and let you know

HHRK I'm still training, it's *very* slow but it works well on a A6000, I've not trai...

Furkan Gözükara SECourses•2/28/25, 3:56 PM

1gb is not big

HRK•2/28/25, 3:57 PM

For a lora ? Would you recommend going rank 256 ? I don't really care about the size or the time it takes to train tbh, as long as the results are there

HHRK For a lora ? Would you recommend going rank 256 ? I don't really care about the ...

Furkan Gözükara SECourses•2/28/25, 4:07 PM

128 usually great

AAIGambino This one I made is slightly better restoration, but not by much!

dupopalot•2/28/25, 8:27 PM

Is this WAN?

AIGambino•3/1/25, 2:13 AM

AAIGambino Click to see attachment

Furkan Gözükara SECourses•3/1/25, 9:06 AM

really good

FFurkan Gözükara SECourses I plan to compare GGUF vs our APP tomorrow hopefully

Timson•3/1/25, 9:36 AM

Are you planning to research gguf for flux dev as well?

TTimson Are you planning to research gguf for flux dev as well?

Furkan Gözükara SECourses•3/1/25, 9:40 AM

you mean compare gguf of flux vs like fp8 or bf16?

FFurkan Gözükara SECourses you mean compare gguf of flux vs like fp8 or bf16?

Timson•3/1/25, 9:43 AM

yep, and also a tutorial on how to use it in swarmUI

TTimson yep, and also a tutorial on how to use it in swarmUI

Furkan Gözükara SECourses•3/1/25, 9:51 AM

i plan to update my downloader and seperate

Furkan Gözükara SECourses•3/1/25, 9:51 AM

flux hunyuan wan etc and including gguf as well

Furkan Gözükara SECourses•3/1/25, 9:51 AM

also found a solution for kaggle

Furkan Gözükara SECourses•3/1/25, 9:51 AM

testing now

Furkan Gözükara SECourses•3/1/25, 9:52 AM

wan 2.1 1.3b model should work fast

FFurkan Gözükara SECourses https://youtu.be/hnAhveNy-8s

Timson•3/1/25, 10:11 AM

awesome tutorial as always, great work!

Timson•3/1/25, 10:16 AM

My only suggestion is using 30 fps video examples, since it's easier to compare with Kling (it always generates 30fps)

TTimson My only suggestion is using 30 fps video examples, since it's easier to compare ...

Furkan Gözükara SECourses•3/1/25, 10:19 AM

i think i will add rifle to increase fps to 30

Furkan Gözükara SECourses•3/1/25, 10:24 AM

i am gonna compile flash attention for swarmui - python 3.12.8

Timson•3/1/25, 10:58 AM

To me bottom line here is - Wan 2.1 14B is not practically usable.
It is on par with Kling 1.6 Pro mode in terms of fidelity, but Kling is both cheaper and much much faster.
50 minutes per generation is not currently usable in AI creative process that requires lots of retries and prompt iteration for any meaningful results.
L40S is at best 1.2$/h currently so it's 1$ per generation (not including time and effort for initial setup) compared with 0.5$ per generation at Kling that completes in 4 minutes.

TTimson To me bottom line here is - Wan 2.1 14B is not practically usable. It is on par ...

Furkan Gözükara SECourses•3/1/25, 11:00 AM

true but difference may come at fine tuning training

Furkan Gözükara SECourses•3/1/25, 11:00 AM

which i plan to cover

TTimson To me bottom line here is - Wan 2.1 14B is not practically usable. It is on par ...