Software Engineering Courses (SECourses)•5mo ago

hey @Dr. Furkan Gözükara, could you give me some advice i have a big checkpoint to train (220 epochs

hey @Dr. Furkan Gözükara, could you give me some advice i have a big checkpoint to train (220 epochs with 300+ images) and it would take 88hours with a L40S. Is there a gpu I can rent what would significantly speed up the training (without going multi gpu, i read here that it impacts the quality of the training) ? Thanks

Sskyrrr hey @Dr. Furkan Gözükara, could you give me some advice i have a big checkpoint ...

Furkan Gözükara SECourses•8/20/25, 8:28 AM

well 220 epoch is too big for 300 images

Furkan Gözükara SECourses•8/20/25, 8:28 AM

it is a single concept right?

Furkan Gözükara SECourses•8/20/25, 8:28 AM

there isnt much you can do to speed up even batch size wont help much

Furkan Gözükara SECourses•8/20/25, 8:28 AM

unless multiple gpu

skyrrrOP•8/20/25, 8:30 AM

@Dr. Furkan Gözükara It's for an editorial style model so the dataset isn't very homogenous... I went with 220 because i'm training for flux krea

skyrrrOP•8/20/25, 8:31 AM

would multiple gpu hurt the quality of the training ?

Sskyrrr @Dr. Furkan Gözükara It's for an editorial style model so the dataset isn't very...

Furkan Gözükara SECourses•8/20/25, 8:35 AM

i see

Sskyrrr would multiple gpu hurt the quality of the training ?

Furkan Gözükara SECourses•8/20/25, 8:35 AM

you can do this

Furkan Gözükara SECourses•8/20/25, 8:35 AM

4 gpus

Furkan Gözükara SECourses•8/20/25, 8:35 AM

batch size 1

Furkan Gözükara SECourses•8/20/25, 8:35 AM

it will minimally impact

Furkan Gözükara SECourses•8/20/25, 8:35 AM

so your total batch size will be 4

Furkan Gözükara SECourses•8/20/25, 8:35 AM

make learning rate = learning rate * 2

skyrrrOP•8/20/25, 8:42 AM

Ok @Dr. Furkan Gözükara thanks a lot. Is there a patreon post where you explain how to set up multiple gpu for runpod ?

Sskyrrr Ok @Dr. Furkan Gözükara thanks a lot. Is there a patreon post where you explain ...

Furkan Gözükara SECourses•8/20/25, 8:46 AM

yep

Furkan Gözükara SECourses•8/20/25, 8:46 AM

https://youtu.be/-uhL2nW7Ddw

YouTubeSECourses

Blazing Fast & Ultra Cheap FLUX LoRA Training on Massed Compute & R...

Unlock the power of FLUX LoRA training, even if you're short on GPUs or looking to boost speed and scale! This comprehensive guide takes you from novice to expert, showing you how to use Kohya GUI for creating top-notch FLUX LoRAs in the cloud. We'll cover everything: maximizing quality, optimizing speed, and finding the best deals. With our exc...

skyrrrOP•8/20/25, 8:48 AM

you're the best

FFurkan Gözükara SECourses make learning rate = learning rate * 2

skyrrrOP•8/20/25, 1:38 PM

Just to make sure, I was using 4e-06, what would that make x2 ?

Sskyrrr Just to make sure, I was using 4e-06, what would that make x2 ?

Furkan Gözükara SECourses•8/20/25, 1:39 PM

4e-06 is 2x of 2e-06

skyrrrOP•8/20/25, 2:23 PM

THanks

skyrrrOP•8/20/25, 2:38 PM

@Dr. Furkan Gözükara there are no 4xGPU config file in the latest kohya zip file

Sskyrrr @Dr. Furkan Gözükara there are no 4xGPU config file in the latest kohya zip file

Furkan Gözükara SECourses•8/20/25, 2:39 PM

not for dreambooth

Furkan Gözükara SECourses•8/20/25, 2:39 PM

it needs 80gb+

Furkan Gözükara SECourses•8/20/25, 2:39 PM

lora has it

Furkan Gözükara SECourses•8/20/25, 2:39 PM

Furkan Gözükara SECourses•8/20/25, 2:40 PM

but you can follow logic of lora and test

skyrrrOP•8/20/25, 2:40 PM

ok will do thanks

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 3:07 PM

my first image was ok, the 2nd one looked like this, why?

1705001-A20full-body20shot20of20ohwx20woman20posing20co-goth-fluxkrea-1-000100.png

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 3:09 PM

hmm, maybe 16bit enabled with the t5xxx_enconly caused it?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 3:52 PM

or this happens, normal image fine, yolo face segment gets pixelated. whats wrong

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 3:54 PM

first pic ok, except for face, next pic with exact same settings and prompt, not even normal pic is fine

AiInfluence•8/20/25, 3:54 PM

youre doing something wrong

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 4:02 PM

obviously. been fine-tuning since september last year and i always use swarmui for grid-comparisons. first time i see this problem. does it have to do with flux krea? am i supposed to use different clips or smth?

AiInfluence•8/20/25, 4:02 PM

what are you settings

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 4:02 PM

same as flux dev

AiInfluence•8/20/25, 4:02 PM

put guidance at 4.5

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 4:03 PM

Prompt: ohwx woman in a 1950s-inspired midi dress with full skirt, small polka dots, three-quarter sleeves, peter pan collar, matching belt at natural waist, kitten heel pumps, delicate bracelet watch, pearl necklace, garden party setting with blooming roses, soft golden hour lighting<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx woman⧉,
Model: goth-fluxkrea-1-000100, Seed: 1991094289, Steps: 40, CFG Scale: 1, Aspect Ratio: Custom, Width: 1280, Height: 1600, Side Length: 1280, Sampler: iPNDM (Improved Pseudo-Numerical methods for Diffusion Models), Scheduler: DDIM Uniform, Flux Guidance Scale: 4, VAE: ae, CLIP-L Model: clip_l, T5-XXL Model: t5xxl_fp16, Preferred DType: Default (16 bit)

AiInfluence•8/20/25, 4:03 PM

also disable scheduler

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/20/25, 4:11 PM

this was the first pic i generated, and it worked, then next pic got pixelated...

1647002-photo20of20ohwx20woman-goth-fluxkrea-1-000100.png

Original message was deleted

AiInfluence•8/20/25, 4:26 PM

see guys , krea is unbeatable

AiInfluence•8/20/25, 4:27 PM

look at that realism

FFurkan Gözükara SECourses Click to see attachment

skyrrrOP•8/20/25, 4:29 PM

I set up 4xL40S but when starting the starting i got this for each of them :

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 22.17 GiB. GPU 0 has a total capacity of 44.40 GiB of which 21.58 GiB is free. Process 271586 has 22.81 GiB memory in use. Of the allocated memory 22.18 GiB is allocated by PyTorch, and 5.70 MiB is reserved by PyTorch but unallocated.

I used the 48GB_GPUz_4x_GPU_Quality_Tier_1_42000MB_10.6_Second_IT_Trains_T5_and_T5_Attention file, then tried with the Quality Tier 3 and still having the same issue. Any clue ?