i didnt get what you mean
i didnt get what you mean
Fif anyone has good wan 2.1 video please send me
Fi can put to tutorial
HHey all, I've played around to try training loras on Wan, I cannot get Diffusion-pipe to work, it never finds the config file, and i've tried DiffSynth, which works but only accepts videos.
I wondered, is there a 3rd alternative ? And as it seems i'm currently stuck with DiffSynth, we do agree the dataset must be videos max 720x1280 of 81 frames at 16fps, right ?
I wondered, is there a 3rd alternative ? And as it seems i'm currently stuck with DiffSynth, we do agree the dataset must be videos max 720x1280 of 81 frames at 16fps, right ?
Fi plan to make a tutorial based on DiffSynth wan training
Fi am not sure about dataset though :d
HYeah 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
HCurrently testing a little training with proper dataset for wan, I will let you know how it goes if that helps
AI'll explain later on currently testing
AI thought you couldn't train yet?
Hnot 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
Hso that's kinda annoying, I can't retrain some of the Hunyuan datasets I have with DiffSynth, which were only images
Hok 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
Going to try different video formats and frame length etc and will let you guys know if it can help
Fawesome please do
Zsigning up for Massed Compute until my ram comes in. getting annoyed of the long load times lol
HSo 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)
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)
Hand 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
Need to test on the 14b model but it's all promising
Z@Dr. Furkan Gözükara im getting some pretty slow speeds with an L40 through Massed compute - do i need to change any settings?
F1.3b could be best bet to train
FL40 is slow
F5090 is surprisingly fast 70 second it for 14b models 
Fyou need to run installer first
Falso download latest zip file
Fdont forget to use downloader as well - model downloade
Fi recorded video extracting
Ftook like 8 hours to prepare
Fnot counting recording just editing
Fi added a low volume (-50) background music i hope you like it
Zwhich card would be the fastest that i can choose on MC? @Dr. Furkan Gözükara
Fmost expensive one
Fbut A100
FH100
Fgood
Zwould the A6000 be faster than the l40?
Fno
Zi feel like my 4090 does the 1.3b model faster than the L40 lol
FA6000 < L40 < L40S and so on
Fyes 4090 faster
F5090 faster than L40S
Fbecause 1.3b model fits into vram
Z56 mins on the l40 to use the 1.4b model is crazy
Fit is wrong
Frtx 3090 takes 8 min
Fso something wrong i need to check later
Fyep completey wrong
FFi can check back later i need to be afk now
Zsounds good thank you
HI'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