Software Engineering Courses (SECourses)•10mo ago

https://youtu.be/hnAhveNy-8s

Wan 2.1 AI Video Model: Ultimate Step-by-Step Tutorial for Windows ...

Alibaba’s new Wan 2.1 text-to-video, video-to-video and image-to-video Open Source AI is unbelievable. In this tutorial I will show how you can install Wan 2.1 all publicly published models into your Windows PC with 1-click installation and use them with the easiest possible way. With the Gradio APP I have developed, you will be able to use Wan ...

Furkan Gözükara SECoursesOP•2/27/25, 10:55 PM

Writing Video chapters and manual fixing of Subtitles in progress

Furkan Gözükara SECoursesOP•2/27/25, 10:55 PM

i got to go now

Furkan Gözükara SECoursesOP•2/27/25, 11:10 PM

I plan to compare GGUF vs our APP tomorrow hopefully

Zaulinator•2/28/25, 12:32 AM

whenever you get a chance Dr, let me know how to speed up the generation time in MC, thanks!

Original message was deleted

hamuz•2/28/25, 4:02 AM

opensource video models still have long way to go

hamuz•2/28/25, 4:02 AM

think about sd1.4 first release

hamuz•2/28/25, 4:02 AM

and we got sd1.5 fine-tunes in just a year

hamuz•2/28/25, 4:03 AM

quality improvements are just incredible between sd1.4 and sd1.5 fine-tunes

hamuz•2/28/25, 4:04 AM

we gonna see better, faster and local video models by end of 2025, im totally sure with that

AIGambino•2/28/25, 5:04 AM

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

Taz•2/28/25, 9:44 AM

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

julius•2/28/25, 11:05 AM

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 SECoursesOP•2/28/25, 3:42 PM

this guy doesnt know

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

it is here

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

how good it works?

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

i plan make a gui for this

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

i plan to update my downloader and seperate

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

flux hunyuan wan etc and including gguf as well

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

also found a solution for kaggle

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

testing now

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

i think i will add rifle to increase fps to 30

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

true but difference may come at fine tuning training

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

now will use like 11 gb

https://youtu.be/hnAhveNy-8s

Similar Threads