Software Engineering Courses (SECourses)•15mo ago

i've done like 12 trainings already testing, works good for sdxl, not much luck for flux still bleed

i've done like 12 trainings already testing, works good for sdxl, not much luck for flux still bleeds

dsienra•10/21/24, 3:41 AM

I'm using the gui, the .tom file is created automatically how can I modify it when I start the traing from the ui?

itOP•10/21/24, 3:42 AM

no no you need to make your own one

itOP•10/21/24, 3:43 AM

the toml thats created automatically is the config one

itOP•10/21/24, 3:45 AM

https://github.com/kohya-ss/sd-scripts/blob/sd3/docs/config_README-en.md

GitHub

sd-scripts/docs/config_README-en.md at sd3 · kohya-ss/sd-scripts

Contribute to kohya-ss/sd-scripts development by creating an account on GitHub.

itOP•10/21/24, 3:45 AM

more info here

itOP•10/21/24, 3:45 AM

itOP•10/21/24, 3:45 AM

put your img and reg folder inside the dataset.toml file

dsienra•10/21/24, 3:49 AM

Ok I think I get it, I can get my last .tom file and add the line to it and then load my custome .tom on "Dataset config file (Optional. Select the toml configuration file to use for the dataset)" on the ui

dsienra•10/21/24, 3:50 AM

I'm right?

itOP•10/21/24, 3:53 AM

no bro you need to make a new txt file with just your img and reg images in it

dsienra•10/21/24, 3:59 AM

a .tom file looks like this

config_lora-20241020-121552.toml2.15KB

itOP•10/21/24, 4:01 AM

thats your config toml

itOP•10/21/24, 4:01 AM

not your dataset toml

dsienra•10/21/24, 4:01 AM

it has mora than my picture images and reg images

dsienra•10/21/24, 4:01 AM

I don use a dataset tom

dsienra•10/21/24, 4:01 AM

I just use the ui

Iit Click to see attachment

itOP•10/21/24, 4:02 AM

Look here where i inserted my dataset.toml file in the gui

itOP•10/21/24, 4:02 AM

this is the only way you can add that reg setting in

dsienra•10/21/24, 4:04 AM

so Differential Output Preservation is not in the UI yet

dsienra•10/21/24, 4:05 AM

itOP•10/21/24, 4:05 AM

brother you need to put it in your dataset.toml file thats the only way, and you load the dataset.toml in the gui

dsienra•10/21/24, 4:06 AM

I get it, but I never used a dataset.tom I never need it

itOP•10/21/24, 4:06 AM

ok idk then

dsienra•10/21/24, 4:11 AM

this is my config, the "Dataset config file (Optional. Select the tom configuration file to use for the dataset)" was always empty on my trainings, I never needed to train anything. may be I'm give it try just to try this new "Differential Output Preservation" option if its the only way

FFurkan Gözükara SECourses This is why you should get minimum 64 GB RAM memory

druhl•10/21/24, 4:28 AM

Will this help with local training only or help speed up the generations as well?

Iit ok idk then

dsienra•10/21/24, 4:30 AM

thanks man, I will investigate the dataset.tom thing I never use it because it was unnecessary

dsienra•10/21/24, 4:32 AM

I just used the img directory with the folders and folder name as captions

Kallamamran•10/21/24, 8:21 AM

Flux is anoying when it comes to upscaling. Squares!

KKallamamran Flux is anoying when it comes to upscaling. Squares!

protector131090•10/21/24, 8:22 AM

that only happens if you upscale 5x. up to 3x never had this problem

protector131090•10/21/24, 8:23 AM

300% zoom (from 2080x3040 res)

protector131090•10/21/24, 8:23 AM

but im using ultimate sd upscaling. dont know what you are using

Kallamamran•10/21/24, 8:26 AM

Hmmm. So incremental upscale to the same resolution should work? :/

Kallamamran•10/21/24, 8:27 AM

I thought it hadto do with final resolution

Kallamamran•10/21/24, 8:28 AM

Trying 2x+2x with the same image

protector131090•10/21/24, 8:35 AM

sd ultimate upscale renders in tiles.

Kallamamran•10/21/24, 8:35 AM

2x + 2x = 4x

protector131090•10/21/24, 8:35 AM

thats why there is no artifacts

Kallamamran•10/21/24, 8:35 AM

Same shit

Kallamamran•10/21/24, 8:35 AM

2048x4096

protector131090•10/21/24, 8:35 AM

try my workflow.

protector131090•10/21/24, 8:35 AM

1 sec

Kallamamran•10/21/24, 8:36 AM

I try shortcuts. I want to get the time down for upscaling and SD takes time

protector131090•10/21/24, 8:36 AM

https://civitai.com/models/698637/comfy-ui-flux-dev-with-upscaler-2-loras-and-adetailer-super-clean

protector131090•10/21/24, 8:38 AM

i dont think you can use shortcuts with flux. lower steps - degrading in quality. fp8 - degrading in quality. In my tests best dev with e4m3f with ultimate sd and --fast argument is the best fastest way

Kallamamran•10/21/24, 8:39 AM

2x+2x is actually WAY worse than 4x in regards to quality

protector131090•10/21/24, 8:40 AM

this image in 50 steps took around 3.5 minutes on 4090. default dtype dev (not fp8)