Software Engineering Courses (SECourses)•10mo ago

For a lora ? Would you recommend going rank 256 ? I don't really care about the size or the time it

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 ...

J2_Nedo•3/1/25, 11:02 AM

true but i prefer local, unrestricted, uncensored creations

Furkan Gözükara SECourses•3/1/25, 11:03 AM

i improved prompt enhance

Timson•3/1/25, 11:03 AM

True, fine tuning and uncensoring are big advantages, but 50 minutes is too much still

Furkan Gözükara SECourses•3/1/25, 11:03 AM

now will use like 11 gb

Furkan Gözükara SECourses•3/1/25, 11:03 AM

about to publish update

Furkan Gözükara SECourses•3/1/25, 11:04 AM

also compiling flash attention for improving our swarmui rtx 5000 series upgrade

Furkan Gözükara SECourses•3/1/25, 11:04 AM

python 3.12

J2_Nedo•3/1/25, 11:04 AM

@Timson i used batch creation the last night, when i woke up the clips where ready

Furkan Gözükara SECourses•3/1/25, 11:05 AM

i plan to add rife fps increase too

FFurkan Gözükara SECourses i plan to add rife fps increase too

Timson•3/1/25, 11:05 AM

is it frame interpolation model?

Timson•3/1/25, 11:05 AM

I bet Kling uses some kind of frame interpolation for perfomance

Timson•3/1/25, 11:07 AM

the problem is - if the model is not trained for reduced fps output it can mess with motion timings

TTimson is it frame interpolation model?

Furkan Gözükara SECourses•3/1/25, 11:07 AM

yes

Furkan Gözükara SECourses•3/1/25, 11:08 AM

https://github.com/hzwer/Practical-RIFE/tree/main?tab=readme-ov-file#model-list

GitHub

GitHub - hzwer/Practical-RIFE: More practical frame interpolation a...

More practical frame interpolation approach. Contribute to hzwer/Practical-RIFE development by creating an account on GitHub.

Furkan Gözükara SECourses•3/1/25, 11:08 AM

i will use this latest one

TTimson the problem is - if the model is not trained for reduced fps output it can mess ...

Furkan Gözükara SECourses•3/1/25, 11:08 AM

well the model already 16 fps

Furkan Gözükara SECourses•3/1/25, 11:08 AM

so we can double and make 32 :d

FFurkan Gözükara SECourses well the model already 16 fps

Timson•3/1/25, 11:12 AM

oh, I didn't realise the measurements are for 80 frames generation (5s at 16 fps). That makes comparsion with Kling even worse as it outputs 150 frames per 4 minute generation

TTimson oh, I didn't realise the measurements are for 80 frames generation (5s at 16 fps...

Furkan Gözükara SECourses•3/1/25, 11:13 AM

ye kling is lead

Furkan Gözükara SECourses•3/1/25, 11:13 AM

but in some cases

Furkan Gözükara SECourses•3/1/25, 11:13 AM

Wan 2.1 made better output i saw comparisons

Furkan Gözükara SECourses•3/1/25, 11:13 AM

by the way when i add double it will be 160 frames for us

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