Software Engineering Courses (SECourses)•16mo ago

Yeah from Accelerate, what about this setting?

�𝔸𝖨ΞӾ Yeah from Accelerate, what about this setting?

GomezBro•9/4/24, 8:40 AM

Yes all good, i loaded his settings bro

GGomezBro Yes all good, i loaded his settings bro

�

𝔸𝖨ΞӾOP•9/4/24, 8:41 AM

Yeah I'm looking at the config template you used, don't know what else to tell ya. I guess wait for the Dr.

GomezBro•9/4/24, 8:41 AM

i relaunched a new instance starting from scratch

�

𝔸𝖨ΞӾOP•9/4/24, 8:41 AM

How much faster is training with 4x GPU?

�

𝔸𝖨ΞӾOP•9/4/24, 8:41 AM

I'm guessing 4x faster lol

GomezBro•9/4/24, 8:42 AM

his config has 0.00005 Learning rate when it loads not 0.0001 like the video shows too, so its not the same

�𝔸𝖨ΞӾ I'm guessing 4x faster lol

GomezBro•9/4/24, 8:42 AM

yup

GomezBro•9/4/24, 8:42 AM

10,000 Images I think in 6 hrs is what im thinking

�

𝔸𝖨ΞӾOP•9/4/24, 8:43 AM

Oh wow damn yeah with dataset that huge it makes sense to go fast

GGomezBro his config has 0.00005 Learning rate when it loads not 0.0001 like the video sho...

�

𝔸𝖨ΞӾOP•9/4/24, 8:47 AM

Well he explained why he changed it in the video here

�𝔸𝖨ΞӾ Well he explained why he changed it in the video here

GomezBro•9/4/24, 8:47 AM

Yes but his loaded settings are different

GomezBro•9/4/24, 8:47 AM

and those are set

�

𝔸𝖨ΞӾOP•9/4/24, 8:47 AM

but not sure why it's not detecting all your GPUs

GomezBro•9/4/24, 8:50 AM

did all fresh same thing, only 1 GPU in use

GomezBro•9/4/24, 9:00 AM

I get an error after a few mins and it errors out, Training has ended

�

𝔸𝖨ΞӾOP•9/4/24, 9:11 AM

Damn bro sorry, Dr should have the antidote ready for you soon

GomezBro•9/4/24, 9:12 AM

there is something off and i cant put my finger on it, because his configs alone should run from loading....

GGomezBro there is something off and i cant put my finger on it, because his configs alone...

�

𝔸𝖨ΞӾOP•9/4/24, 9:13 AM

You've successfully trained several Loras with Kohya already, right? I don't remember

�𝔸𝖨ΞӾ You've successfully trained several Loras with Kohya already, right? I don't rem...

GomezBro•9/4/24, 9:14 AM

Yeah on my machines at home i have no issues, perfect setups

GomezBro•9/4/24, 9:14 AM

I have A6000 at home

GomezBro•9/4/24, 9:14 AM

and 3 3090 Ti 24 GB

�

𝔸𝖨ΞӾOP•9/4/24, 9:14 AM

Daaayyyuuum

GomezBro•9/4/24, 9:14 AM

I did 1000 Images Lora and its sick, but I want to do 10000 on 4 GPU

GomezBro•9/4/24, 9:15 AM

so it doesnt take long, it should do 10,000 steps in 6hrs

GomezBro•9/4/24, 9:15 AM

and each can be 1 epoch

�

𝔸𝖨ΞӾOP•9/4/24, 9:16 AM

Ok I'm communicating with a certified boss playa then lmao so the issue is isolated to the cloud GPU setup then hmm

�𝔸𝖨ΞӾ Ok I'm communicating with a certified boss playa then lmao so the issue is isola...

GomezBro•9/4/24, 9:16 AM

yeah, the config, something is off and crashing on the 4 GPU config or some other issue

МИШАНЯ•9/4/24, 9:45 AM

Greetings 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

Neuromasters•9/4/24, 10:33 AM

And why do we need so many steps, a range of 2k to 4k is not enough for styling on items?

Neuromasters•9/4/24, 10:37 AM

we 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?

Ffoobax Any tips on getting more realistic results? You've replicated the face very well...

Furkan Gözükara SECourses•9/4/24, 12:08 PM

use more and better training images

Furkan Gözükara SECourses•9/4/24, 12:08 PM

it really works with flux

TTrumpYouAll Dr Gözükara, can you shed light on choices of 'Samplers" and their respective sc...

Furkan Gözükara SECourses•9/4/24, 12:08 PM

i use unipc

Furkan Gözükara SECourses•9/4/24, 12:08 PM

i did some grid testing of all and i think unipc is best

JJackie I have a question about training LoRA. My LoRA files usually have a very large s...

Furkan Gözükara SECourses•9/4/24, 12:09 PM

we save as float : double size. you can save as fp16. our loras are 128 rank

Original message was deleted

Furkan Gözükara SECourses•9/4/24, 12:10 PM

you are supposed to divide your epoch to / number of GPUs

Furkan Gözükara SECourses•9/4/24, 12:10 PM

i told this in video

NNeuromasters And why do we need so many steps, a range of 2k to 4k is not enough for styling ...

Furkan Gözükara SECourses•9/4/24, 12:10 PM

it is experimental

Furkan Gözükara SECourses•9/4/24, 12:10 PM

once i analyzed results we will see

Furkan Gözükara SECourses•9/4/24, 12:10 PM

which steps yielding good :d

Furkan Gözükara SECourses•9/4/24, 12:10 PM

let me upload them to hugging face

Furkan Gözükara SECourses•9/4/24, 12:10 PM

you can also test

FFurkan Gözükara SECourses let me upload them to hugging face

Neuromasters•9/4/24, 12:14 PM

I'm just asking, I inquired with the SDXL models coach lye a long time ago. they did their experiments and gave their metrics which I adhered to. Characters are in the 1200-1800-2000 range.
Objects 1500-2500.
Styles 1200-3000 depending on elements and creativity.
Architectural objects up to 4000 with increased dim and alpfa.
Now I wonder what the figures are for Flux

NNeuromasters I'm just asking, I inquired with the SDXL models coach lye a long time ago. they...

Furkan Gözükara SECourses•9/4/24, 12:16 PM

i am uploading models

Furkan Gözükara SECourses•9/4/24, 12:16 PM

from 25 to 450 epoch

Furkan Gözükara SECourses•9/4/24, 12:16 PM

114 images

Furkan Gözükara SECourses•9/4/24, 12:16 PM

i will also test hopefully

FFurkan Gözükara SECourses from 25 to 450 epoch

Neuromasters•9/4/24, 12:17 PM

the base step is 1?