and batch is at 1 not 4 too
and batch is at 1 not 4 too
�Oh ok gotcha, I'm barely watching the video rn
Gtheres just a couple conflicts from video/4xgpu gpu configs that you load
Gis a bit grey
�This is from the video, you sure this was configured correctly?
Gthats under accelerated right?
GI checked it against his video and loaded his settings bro
Gthose 4X gpu settings are set
Gthat had to be set and was
�Yeah from Accelerate, what about this setting?
GYes all good, i loaded his settings bro
�Yeah I'm looking at the config template you used, don't know what else to tell ya. I guess wait for the Dr.
Gi relaunched a new instance starting from scratch
�How much faster is training with 4x GPU?
�I'm guessing 4x faster lol
Ghis config has 0.00005 Learning rate when it loads not 0.0001 like the video shows too, so its not the same
Gyup
G10,000 Images I think in 6 hrs is what im thinking
�Oh wow damn yeah with dataset that huge it makes sense to go fast
�Well he explained why he changed it in the video here
GYes but his loaded settings are different
Gand those are set
�but not sure why it's not detecting all your GPUs
Gdid all fresh same thing, only 1 GPU in use
GI get an error after a few mins and it errors out, Training has ended
�Damn bro sorry, Dr should have the antidote ready for you soon 
Gthere is something off and i cant put my finger on it, because his configs alone should run from loading....
�You've successfully trained several Loras with Kohya already, right? I don't remember
GYeah on my machines at home i have no issues, perfect setups
GI have A6000 at home
Gand 3 3090 Ti 24 GB
�Daaayyyuuum
GI did 1000 Images Lora and its sick, but I want to do 10000 on 4 GPU
Gso it doesnt take long, it should do 10,000 steps in 6hrs
Gand each can be 1 epoch
�Ok I'm communicating with a certified boss playa then lmao so the issue is isolated to the cloud GPU setup then hmm
Gyeah, the config, something is off and crashing on the 4 GPU config or some other issue
МGreetings to all AI and SD specialists.
I came across the following question. In SD is it realistic to make ONE image from different positions ? For example: Girl: Front, back, side view. At the same proportions, appearance that would not change, I want to create a 2d game, for this I need the same sprites from different positions
I came across the following question. In SD is it realistic to make ONE image from different positions ? For example: Girl: Front, back, side view. At the same proportions, appearance that would not change, I want to create a 2d game, for this I need the same sprites from different positions
NAnd why do we need so many steps, a range of 2k to 4k is not enough for styling on items?
Nwe have a token to bake owix, it is empty, we have described the interactions with it. We gave a prototype that it is an icon object, etc.
steps for retraining other objects in the model memory are missing. Then why such a huge number of steps?
steps for retraining other objects in the model memory are missing. Then why such a huge number of steps?
Fuse more and better training images
Fit really works with flux
Fi use unipc
Fi did some grid testing of all and i think unipc is best
Fwe save as float : double size. you can save as fp16. our loras are 128 rank
Fyou are supposed to divide your epoch to / number of GPUs
Fi told this in video
Fit is experimental
Fonce i analyzed results we will see
Fwhich steps yielding good :d
Flet me upload them to hugging face
