Software Engineering Courses (SECourses)•10mo ago

i am not sure about dataset though :d

HRK•2/27/25, 2:54 PM

Yeah it's not super clear, but I cannot make image work in training, the data process step returns this : "AttributeError: 'LegacyReader' object has no attribute 'count_frames'", I'm not a tech wizard but to me it means no images and only videos ahah

FFurkan Gözükara SECourses i am not sure about dataset though :d

HRK•2/27/25, 2:55 PM

Currently testing a little training with proper dataset for wan, I will let you know how it goes if that helps

FFurkan Gözükara SECourses i didnt get what you mean

AiInfluence•2/27/25, 3:00 PM

I'll explain later on currently testing

HHRK Hey all, I've played around to try training loras on Wan, I cannot get Diffusion...

AiInfluence•2/27/25, 3:00 PM

I thought you couldn't train yet?

AAiInfluence I thought you couldn't train yet?

HRK•2/27/25, 3:05 PM

not on diffusionpipe, the config file is apparently never found, but I can on diffsynth, but I can't train with images on diffsynth, only videos

HRK•2/27/25, 3:07 PM

so that's kinda annoying, I can't retrain some of the Hunyuan datasets I have with DiffSynth, which were only images

HRK•2/27/25, 3:16 PM

ok I'm super super dumb I forgot to remove a " / " in the config file which made the whole thing go mad for the config.json file, diffusionpipe works

Going to try different video formats and frame length etc and will let you guys know if it can help

HHRK ok I'm super super dumb I forgot to remove a " / " in the config file which made...

Furkan Gözükara SECoursesOP•2/27/25, 3:34 PM

awesome please do

Zaulinator•2/27/25, 4:21 PM

signing up for Massed Compute until my ram comes in. getting annoyed of the long load times lol

FFurkan Gözükara SECourses awesome please do

HRK•2/27/25, 4:26 PM

So i've barely scraped the surface but on diffusion-pipe I've tried simple videos at 16fps. Following the num_frames from the empty embeds node from Kijai, I've used 16n+1 number of frames, (17, 33, 49, 65, 81, 129). I added all these in the frame_bucket part of the dataset.toml to be safe, so it looks like that : frame_buckets = [1, 17, 33, 49, 65, 81, 129] (1 is for Images I guess).

Basically everything got processed but not the 129 frames clip, i suppose it's too long.

Running on linux on a 4090, I can train the lora at rank 128, resolution 704, and it uses around 19gb of the gpu. All that training on the wan 1.3b model. I've tried 1024 training but it ooms instantly. (knowing my system was taking 400mb, I don't think it's possible to train 1024 at all on a 4090)

HRK•2/27/25, 4:41 PM

and with the exact same parameters, when using only images in the dataset it seems faster than hunyuan and uses only 14gb vram

Need to test on the 14b model but it's all promising

Zaulinator•2/27/25, 4:55 PM

@Dr. Furkan Gözükara im getting some pretty slow speeds with an L40 through Massed compute - do i need to change any settings?

ZZaulinator signing up for Massed Compute until my ram comes in. getting annoyed of the long...

Furkan Gözükara SECoursesOP•2/27/25, 5:31 PM

HHRK So i've barely scraped the surface but on diffusion-pipe I've tried simple video...

Furkan Gözükara SECoursesOP•2/27/25, 5:31 PM

1.3b could be best bet to train

ZZaulinator @Dr. Furkan Gözükara im getting some pretty slow speeds with an L40 through Mas...

Furkan Gözükara SECoursesOP•2/27/25, 5:31 PM

L40 is slow

Furkan Gözükara SECoursesOP•2/27/25, 5:31 PM

5090 is surprisingly fast 70 second it for 14b models

Original message was deleted

Furkan Gözükara SECoursesOP•2/27/25, 5:32 PM

you need to run installer first

Furkan Gözükara SECoursesOP•2/27/25, 5:32 PM

also download latest zip file

Furkan Gözükara SECoursesOP•2/27/25, 5:32 PM

dont forget to use downloader as well - model downloade

Furkan Gözükara SECoursesOP•2/27/25, 5:32 PM

i recorded video extracting

Furkan Gözükara SECoursesOP•2/27/25, 5:33 PM

took like 8 hours to prepare

Furkan Gözükara SECoursesOP•2/27/25, 5:33 PM

not counting recording just editing

Furkan Gözükara SECoursesOP•2/27/25, 5:33 PM

i added a low volume (-50) background music i hope you like it

FFurkan Gözükara SECourses L40 is slow

Zaulinator•2/27/25, 5:37 PM

which card would be the fastest that i can choose on MC? @Dr. Furkan Gözükara

ZZaulinator which card would be the fastest that i can choose on MC? @Dr. Furkan Gözükara

Furkan Gözükara SECoursesOP•2/27/25, 5:39 PM

most expensive one

Furkan Gözükara SECoursesOP•2/27/25, 5:39 PM

but A100

Furkan Gözükara SECoursesOP•2/27/25, 5:39 PM

H100

Furkan Gözükara SECoursesOP•2/27/25, 5:39 PM

good

FFurkan Gözükara SECourses most expensive one 😄

Zaulinator•2/27/25, 5:39 PM

would the A6000 be faster than the l40?

ZZaulinator would the A6000 be faster than the l40?

Furkan Gözükara SECoursesOP•2/27/25, 5:39 PM

Zaulinator•2/27/25, 5:39 PM

i feel like my 4090 does the 1.3b model faster than the L40 lol

Furkan Gözükara SECoursesOP•2/27/25, 5:40 PM

A6000 < L40 < L40S and so on

Furkan Gözükara SECoursesOP•2/27/25, 5:40 PM

yes 4090 faster

Furkan Gözükara SECoursesOP•2/27/25, 5:40 PM

5090 faster than L40S

Furkan Gözükara SECoursesOP•2/27/25, 5:40 PM

because 1.3b model fits into vram

Zaulinator•2/27/25, 5:40 PM

56 mins on the l40 to use the 1.4b model is crazy

ZZaulinator 56 mins on the l40 to use the 1.4b model is crazy

Furkan Gözükara SECoursesOP•2/27/25, 5:43 PM

it is wrong

Furkan Gözükara SECoursesOP•2/27/25, 5:43 PM

rtx 3090 takes 8 min

Furkan Gözükara SECoursesOP•2/27/25, 5:43 PM

so something wrong i need to check later

FFurkan Gözükara SECourses so something wrong i need to check later 😄

Zaulinator•2/27/25, 5:44 PM

ZZaulinator Click to see attachment

Furkan Gözükara SECoursesOP•2/27/25, 5:44 PM

yep completey wrong

Furkan Gözükara SECoursesOP•2/27/25, 5:44 PM

i can check back later i need to be afk now

FFurkan Gözükara SECourses i can check back later i need to be afk now

Zaulinator•2/27/25, 5:45 PM

sounds good thank you

FFurkan Gözükara SECourses 1.3b could be best bet to train

HRK•2/27/25, 6:04 PM

I've just tried a lora I trained for 16 epoch with ~250 Images, the results are better than hunyuan no contest, it's very early in the training and yet it's sharp and looks quite good, i will share things when I get to a point where the training is perfect

HRK•2/27/25, 6:05 PM

while I would usualy train a bunch of videos first at res 512 then switch to images at res 1024, here only images at 704 and it's alright

uisato•2/27/25, 6:47 PM

Liminal Found Footage - [Flux Experiment]

More experiments, and project files, through: https://linktr.ee/uisato

https://www.youtube.com/shorts/dP3yCiqkHzA

Zaulinator•2/27/25, 6:56 PM

yeah tried every combo of settings in MC with the l40, still like 10x slower for the 1.3b model than my 4090, so something is funky is going on.

V3ldrin•2/27/25, 9:00 PM

Hey, I was wondering why I get an "RuntimeError: CUDA error: out of memory" in kohya when using your "lora training best flux configs" with 16GB and less. (have a 24 GB card, and the 24 GB configs do work! ... )

FFurkan Gözükara SECourses compared to which gpu?

sergio_gonzo•2/27/25, 10:36 PM

compared to 4070 ti super

i am not sure about dataset though :d

Similar Threads