my unified downloader also downloads it
my unified downloader also downloads it
Fhow good it works?
Fi plan make a gui for this
Fby the way sd 3.5 large training getting improved. now we gonna have trained clip l and clip g into embedded file
HI'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
F1gb is not big 
HFor 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
F128 usually great
DIs this WAN?
AFreally good
TAre you planning to research gguf for flux dev as well?
Fyou mean compare gguf of flux vs like fp8 or bf16?
Tyep, and also a tutorial on how to use it in swarmUI
Fi plan to update my downloader and seperate
Fflux hunyuan wan etc and including gguf as well
Falso found a solution for kaggle
Ftesting now
Fwan 2.1 1.3b model should work fast
Tawesome tutorial as always, great work!
TMy only suggestion is using 30 fps video examples, since it's easier to compare with Kling (it always generates 30fps)
Fi think i will add rifle to increase fps to 30
Fi am gonna compile flash attention for swarmui - python 3.12.8
TTo 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.
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.
Ftrue but difference may come at fine tuning training
Fwhich i plan to cover
Jtrue but i prefer local, unrestricted, uncensored creations 
Fi improved prompt enhance
TTrue, fine tuning and uncensoring are big advantages, but 50 minutes is too much still
Fnow will use like 11 gb
Fabout to publish update
Falso compiling flash attention for improving our swarmui rtx 5000 series upgrade
Fpython 3.12
J@Timson i used batch creation the last night, when i woke up the clips where ready
Fi plan to add rife fps increase too
Tis it frame interpolation model?
TI bet Kling uses some kind of frame interpolation for perfomance
Tthe problem is - if the model is not trained for reduced fps output it can mess with motion timings
Fyes
FGitHub
More practical frame interpolation approach. Contribute to hzwer/Practical-RIFE development by creating an account on GitHub.
Fi will use this latest one
Fwell the model already 16 fps
Fso we can double and make 32 :d
Toh, 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
Fye kling is lead
Fbut in some cases
FWan 2.1 made better output i saw comparisons
Fby the way when i add double it will be 160 frames for us
A@Dr. Furkan Gözükara this is exactly what. I meant on Patreon chat
Ais it even worth the research time when img2 vid exists?
Ji saw people train Lora for hunyuan, i hope this comes to Wan too
