Software Engineering Courses (SECourses)•16mo ago

false alarm sorry

dsienraOP•9/18/24, 6:07 PM

my dataset es low ress

dsienraOP•9/18/24, 6:07 PM

thats why

dsienraOP•9/18/24, 6:08 PM

I have bukets enable

dxqb•9/18/24, 6:08 PM

ah yeah, no upscale on kohya. resolution makes a major difference on flux

dxqb•9/18/24, 6:12 PM

what resolution was it actually?

dsienraOP•9/18/24, 6:14 PM

some 768 some 512

dsienraOP•9/18/24, 6:15 PM

not all 1:1 diferent aspect ratios

FFurkan Gözükara SECourses please ask there

Marrvelous•9/18/24, 6:18 PM

Thanks. I posted there but there is a long conversation already happening so probably missed. I'll try posting again later

ZZono50 @Dr. Furkan Gözükara on supir app, if guidance scale needs to be 2, where do i ...

Furkan Gözükara SECourses•9/18/24, 8:10 PM

yes

Ddsienra Click to see attachment

Furkan Gözükara SECourses•9/18/24, 8:10 PM

good speed

Ddsienra Click to see attachment

Furkan Gözükara SECourses•9/18/24, 8:10 PM

nope

Ddsienra false alarm sorry

Furkan Gözükara SECourses•9/18/24, 8:11 PM

yep i was gonna tell this :d

TheJayHole•9/18/24, 8:39 PM

New supporter here. Wanted to say thanks. Your info is great. Couple of questions for you. I'm doing LoRA training on Flux Really, really slow on my 3090 (Windows with latest torch 2.5 speedup), 51 images in LoRA. This may have been answered somewhere, but previously with this SDXL LoRA I would do 32 repeats, 51 images, 10 epoch (16320 steps). This was a little high admittedly. The config you had would be 10200 steps Are epochs better than repeats with FLUX1.dev? Does this lower the number required? Just looking to understand that a little better. As for the slowness I'm 4735 steps and running on 1.5-2 days. I did update torch. Pretty much the same speed for me. Might have to try on RunPod. Any thoughts/advice are appreciated.

TTheJayHole New supporter here. Wanted to say thanks. Your info is great. Couple of quest...

Furkan Gözükara SECourses•9/18/24, 8:53 PM

welcome

Furkan Gözükara SECourses•9/18/24, 8:53 PM

what is your step speed?

Furkan Gözükara SECourses•9/18/24, 8:53 PM

you can use Rank_4_17250MB_4_85_Second_IT.json it is pretty decent speed

Furkan Gözükara SECourses•9/18/24, 8:53 PM

200 epoch may be a lot for 51 image

Furkan Gözükara SECourses•9/18/24, 8:53 PM

but you can compare checkpoints

xlads•9/18/24, 9:25 PM

Is nsfw possible with flux lora?

Xxlads Is nsfw possible with flux lora?

Furkan Gözükara SECourses•9/18/24, 9:29 PM

i dont know i really dont into nsfw

Furkan Gözükara SECourses•9/18/24, 9:29 PM

we also avoid discussing it here

xlads•9/18/24, 9:36 PM

Okay! Asking for someone else

dxqb•9/18/24, 9:39 PM

I've just trained a very bad dataset on flux. Blurry, mixed resolution down to 128px. For SDXL, I have carefully prepared it using SUPIR and more. With flux, it just works without preparation. This model is insane

Ddxqb I've just trained a very bad dataset on flux. Blurry, mixed resolution down to 1...

Furkan Gözükara SECourses•9/18/24, 10:00 PM

so true

Furkan Gözükara SECourses•9/18/24, 10:00 PM

with flux you can only get better

Furkan Gözükara SECourses•9/18/24, 10:00 PM

i plan to do general fine tuning with our 20k images but i need at least 2x A100 like 1 week

Furkan Gözükara SECourses•9/18/24, 10:01 PM

48gb fine tuning fails at multi gpu

dxqb•9/18/24, 10:05 PM

what's the goal?

Ddxqb what's the goal?

Furkan Gözükara SECourses•9/18/24, 10:09 PM

that model may work better for training a person afterwards

Furkan Gözükara SECourses•9/18/24, 10:09 PM

not sure though a good experiment

dxqb•9/18/24, 10:12 PM

ah, yeah might be a good model. it can be trained towards realism very easily, but without any lora it looks very AI generated

dxqb•9/18/24, 10:12 PM

probably was trained on generated data

Furkan Gözükara SECourses•9/19/24, 12:13 AM

started a fine tuning training for a client with 100 images

Furkan Gözükara SECourses•9/19/24, 12:13 AM

very unique case

Furkan Gözükara SECourses•9/19/24, 12:13 AM

Furkan Gözükara SECourses•9/19/24, 12:13 AM

lets see what we will get

Furkan Gözükara SECourses•9/19/24, 12:14 AM

it sucks that fine tuning doesnt fit into 48 gb with multi gpu

Furkan Gözükara SECourses•9/19/24, 1:07 AM

just published an amazing article : https://www.patreon.com/posts/112335162?pr=true

Patreon

How to Extract LoRA from FLUX Fine Tuning / DreamBooth Training Ful...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

Neo•9/19/24, 2:24 AM

oooooooh thank you doc

FFurkan Gözükara SECourses started a fine tuning training for a client with 100 images

Neo•9/19/24, 2:26 AM

Thanks, it will be useful to me as well, as discussed before with you I often have datasets in that range (100-250imgs). Curious as to the concept category being trained, is this for a person or style, etc?

NNeo Thanks, it will be useful to me as well, as discussed before with you I often ha...

Furkan Gözükara SECourses•9/19/24, 2:54 AM

Person

Furkan Gözükara SECourses•9/19/24, 2:54 AM

I plan to general fine tune with 20400 images

Furkan Gözükara SECourses•9/19/24, 2:54 AM

Captioning right now

Furkan Gözükara SECourses•9/19/24, 2:54 AM

But it will be on single gpu :/

Darth Teethius•9/19/24, 4:00 AM

Has anyobe figured out the specific block training in kohya?

FFurkan Gözükara SECourses you can use Rank_4_17250MB_4_85_Second_IT.json it is pretty decent speed

TheJayHole•9/19/24, 4:08 AM

TYVM

FFurkan Gözükara SECourses just published an amazing article : https://www.patreon.com/posts/112335162?pr=t...

dxqb•9/19/24, 5:35 AM

do kohya merge tools already work with flux? I'd like to train on a checkpoint that was merged with a Lora

dxqb•9/19/24, 5:36 AM

and, do you already have any insights on DoRA? the claim is that it closes the gap between Lora and Finetune. Is it correct? OT has implemented Dora recently, I think kohya had it for longer

Kallamamran•9/19/24, 5:59 AM

What does the numbers under the image in Forge mean?
A: 18.65 GB, R: 20.05 GB, Sys: 21.7/24 GB (90.5%)

R=RAM? Sys=VRAM? A=??

liorftw•9/19/24, 7:06 AM

Hey I saw there is new things that he posted, the last thing I did is train a lora with my data, what's now the best thing? I saw there is lora extraction what does it mean?

false alarm sorry

Similar Threads