Is this the recent clip_l trained one?
Is this the recent clip_l trained one?
SAnd are you using captions for these?
Fonly ohwx man atm
Fyes clip 5e-05
PHave there been any updates that make Supir useable for video?
When using settings 1,1,2,1, a batch upscale of 8700 images with the original resolution at 960x576 and using Upscale Size 1 it still takes nearly 16 hours using a 4090...
If this is how Supir is supposed to function it is not a viable option for video.
When using settings 1,1,2,1, a batch upscale of 8700 images with the original resolution at 960x576 and using Upscale Size 1 it still takes nearly 16 hours using a 4090...
If this is how Supir is supposed to function it is not a viable option for video.
AI send you message in private, I think to be important especially for lora training
Fthey are working on video model
PI am batch upscaling individual images that comprise video, have there been any updates that speed up Supir for batch processing?
PI just don't understand why it is so painfully slow...
M@Dr. Furkan Gözükara i think your kohya installer is broken.
im now getting windows os errors, and gui itself doesn't want to open.
im now getting windows os errors, and gui itself doesn't want to open.
Fcan you tell me error
Fsadly not but i feel you
Fi wonder if pytorch 2.5 can speed it up
PIt shows I am only using 7.2gb of my GPU ram
Palso states gradio 4.29.0 is available and it is using 4.19 during startup selection
Mwell im not getting the win os error now because the gui.bat is not doing anything anymore
Mits just black screen
Mi was getting permission errors as well
Fare you on windows 11?
Myessir
Fwindows 11 sucks
Frun gui.bat inside kohya ss folder and tell me waht you see
Frestart computer and try again too
Mi know i had to upgrade though, windows 10 was giving so much issues, random pc shutdowns, slowdowns etc. plus windows 10 soon wont be supported anymore
Malready tried i was getting just a blank terminal
Mbut im doing system restore right now back to 4 days ago
Fdid you get any error at install?
Fwindows 11 has problematic permissions with python apps :/
Myeah i was getting some windows os error
Fyou need to fix it
Fit is permission related error
M
Fi still use windows 10 
but i also get random shutdowns and bluescreens sometimes :/
but i also get random shutdowns and bluescreens sometimes :/
PHey there, thank you for the plethora of information, I'm learning a ton from all the work you're putting into this.
I'm trying to understand how Koyha is calculating the total optimization steps when I run my LORA training...
I have 30 images set to 10 repeats, so that part is clear, but it somehow seems to be automatically configuring things to hit 1600 steps.
The Epoch setting in Koyha doesn't seem to actually change anything it looks like, so where does it come up with 6 epochs?
I'm trying to understand how Koyha is calculating the total optimization steps when I run my LORA training...
I have 30 images set to 10 repeats, so that part is clear, but it somehow seems to be automatically configuring things to hit 1600 steps.
running training / 学習開始
num train images * repeats / 学習画像の数×繰り返し回数: 300
num reg images / 正則化画像の数: 0
num batches per epoch / 1epochのバッチ数: 300
num epochs / epoch数: 6
batch size per device / バッチサイズ: 1
gradient accumulation steps / 勾配を合計するステップ数 = 1
total optimization steps / 学習ステップ数: 1600The Epoch setting in Koyha doesn't seem to actually change anything it looks like, so where does it come up with 6 epochs?
DI'm doing new tests with TE training multiple people but it seems that still bleeds to the class, may be for flux you have to remove the class completely to avoid bleeding the people you are training between each other, and the bleed to the entire class "man" for example, may be changing the captions from "ohwx man" to only "ohwx" will do the magic, flux already knows what a man is and is not necessary to add it to the captions, In sdxl this wasn't a problem for me. the model learned that "name1 man" was different from "name2 man" in flux I didn't find a solution yet, I will try removing the class in my next training.
Fi explained in the video
Fyou must be setting somwhere epoch
For max epoch
Fstill ediitng video to publish
Fyes told this in video
Fdont use class
Fit will work better
Fi tested just ohwx man and since its internally already captioning
Fit works
Fstill will likely to bleed but lesser
Pahh, ok. I'll look into it more, thank you.
Fonce video published please watch fully
Pwill do!
Doff course I'm waiting... thanks
Dthe speed on WSL is inconsistent yesterday speed
Dtoday speed
