Software Engineering Courses (SECourses)•7mo ago

probably extraction logic is broken not accurate

Furkan Gözükara SECoursesOP•5/31/25, 12:46 PM

for flux

FFurkan Gözükara SECourses probably extraction logic is broken not accurate

AiInfluence•5/31/25, 1:21 PM

then why you say you reccomend it

Original message was deleted

AiInfluence•5/31/25, 1:21 PM

yeah only problem is the line artifacts that come up

AiInfluence•5/31/25, 1:22 PM

swarm

AiInfluence•5/31/25, 1:22 PM

i put 2k res

AiInfluence•5/31/25, 1:22 PM

the lines only appear when using a lora

AiInfluence•5/31/25, 1:25 PM

yeah

AiInfluence•5/31/25, 1:25 PM

i found supir

AiInfluence•5/31/25, 1:25 PM

so it gets the job done

AiInfluence•5/31/25, 1:25 PM

but its incconvienti you know

AiInfluence•5/31/25, 1:25 PM

db fixes those artifacts

AiInfluence•5/31/25, 1:25 PM

just amazing

AiInfluence•5/31/25, 1:26 PM

i never tried to upscale in swarm

AiInfluence•5/31/25, 1:26 PM

can you guide me?

AiInfluence•5/31/25, 1:26 PM

sure

AAiInfluence then why you say you reccomend it

Furkan Gözükara SECoursesOP•5/31/25, 1:51 PM

i dont recommend for flux anymore

Furkan Gözükara SECoursesOP•5/31/25, 1:51 PM

after i did some more testing and usage

Furkan Gözükara SECoursesOP•5/31/25, 1:52 PM

but it has 1 advantage

Furkan Gözükara SECoursesOP•5/31/25, 1:52 PM

if you upscale to 2x, extract lora works better than lora since lora produces lines

Elli_Oaken•6/2/25, 12:28 PM

Hi guys! I've been training FLUX LoRAs using Kohya on Windows with 4 RTX GPUs. About a month ago, everything worked fine, but now I keep getting crashes related to torch.distributed.get_world_size() when using Accelerate.

Has something changed recently in PyTorch or Accelerate that breaks multi-GPU training on Windows?
Is there any recommended workaround for this?

EElli_Oaken Hi guys! I've been training FLUX LoRAs using Kohya on Windows with 4 RTX GPUs. A...

Furkan Gözükara SECoursesOP•6/2/25, 2:52 PM

yes you need like

Furkan Gözükara SECoursesOP•6/2/25, 2:52 PM

SXM or NVLINK machines

Original message was deleted

AiInfluence•6/6/25, 10:12 AM

Hi Rex , when did he state that?

Original message was deleted

AiInfluence•6/6/25, 10:15 AM

so whats the workaround

AiInfluence•6/6/25, 10:15 AM

since lines dont go away, maybe train on a custom checkpoint?

Original message was deleted

AiInfluence•6/6/25, 10:17 AM

on 2k res with loras?

AiInfluence•6/6/25, 10:17 AM

how

AiInfluence•6/6/25, 10:18 AM

1024x1024

AiInfluence•6/6/25, 10:18 AM

but i use community loras too so idk theirs

AiInfluence•6/6/25, 10:18 AM

such as?

AiInfluence•6/6/25, 10:19 AM

can you dm me an example dataset?

AiInfluence•6/6/25, 10:20 AM

so for 1024x1024 i shouldnt go higher?

AiInfluence•6/6/25, 10:20 AM

with dreambooth im not getting lines tho and it works perfect

AiInfluence•6/6/25, 10:21 AM

yeah

AiInfluence•6/6/25, 10:21 AM

and i tested in extreme resolutions

AiInfluence•6/6/25, 10:21 AM

and results where outstanding

AiInfluence•6/6/25, 10:21 AM

that's why your comment on not training the full model anymore confused me

AiInfluence•6/6/25, 10:21 AM

i know its overkill for 20images to train the whole network but if it works it works

AiInfluence•6/6/25, 10:22 AM

until a solution arives

Original message was deleted

AiInfluence•6/6/25, 10:22 AM

on replicate?

AiInfluence•6/6/25, 10:22 AM

ostris?

AiInfluence•6/6/25, 10:23 AM

oh you dont use kohya

AiInfluence•6/6/25, 10:23 AM

how so

Original message was deleted

AiInfluence•6/6/25, 10:23 AM

should i train that ? what do you think

AiInfluence•6/6/25, 10:24 AM

what does wip stand for

Original message was deleted

AiInfluence•6/6/25, 10:26 AM

db of course

AiInfluence•6/6/25, 10:27 AM

in visual quality more

AiInfluence•6/6/25, 10:27 AM

i trained a subject with 5 images

AiInfluence•6/6/25, 10:27 AM

very limited dataset

AiInfluence•6/6/25, 10:27 AM

and results where amazingf

probably extraction logic is broken not accurate

Similar Threads