Software Engineering Courses (SECourses)•15mo ago

<@205854764540362752> Have you looked into flux rapid lora training? Both fal and civit.ai offer i

@Furkan Gözükara SECourses Have you looked into flux rapid lora training? Both fal and civit.ai offer it.

https://education.civitai.com/quickstart-guide-to-flux-1/#rapid-flux-training

VVigilence <@205854764540362752> Have you looked into flux rapid lora training? Both fal ...

Furkan Gözükara SECourses•10/15/24, 9:08 PM

i checked civitiai

Furkan Gözükara SECourses•10/15/24, 9:08 PM

as expected it was inferior

Furkan Gözükara SECourses•10/15/24, 9:08 PM

i didnt pay for fal

Furkan Gözükara SECourses•10/15/24, 9:08 PM

i know it will be too

VigilenceOP•10/15/24, 9:10 PM

I think if you make your own version you would get lots of begginers to payfor the config. I know I would have signed up for it back then.

wardensc2•10/16/24, 4:32 AM

Hi @Furkan Gözükara SECourses I use your Batch_Size_7_48GB_GPU_46250MB_28.9_second_it_Tier_1. In the config I only see 1 GPU not multi GPU, so you only increase the batch file to 7 right. What about the time and quality if I use your batch 7, does it save me 7 times less than your batch 1 config and the quality is the same ?

Constipated•10/16/24, 6:01 AM

@Furkan Gözükara SECourses saw your Flux 15 images vs 256 analysis and it was very insightful. Obviously 256 images may not be practical in many cases and would take a long time on Batch 1. Let's say I have the more realistic option of using 30 or 40 images. Do you think that is better than 15?

Constipated•10/16/24, 6:02 AM

Assuming all are high quality images

Wwardensc2 Hi <@205854764540362752> I use your Batch_Size_7_48GB_GPU_46250MB_28.9_second_it...

Furkan Gözükara SECourses•10/16/24, 9:50 AM

if you get 7 GPU SXM machine with A100

Furkan Gözükara SECourses•10/16/24, 9:50 AM

it will be like 7 times faster

Furkan Gözükara SECourses•10/16/24, 9:50 AM

and you can use same learning rate

CConstipated <@205854764540362752> saw your Flux 15 images vs 256 analysis and it was very in...

Furkan Gözükara SECourses•10/16/24, 9:50 AM

definitely better

Furkan Gözükara SECourses•10/16/24, 9:50 AM

15 images underperform

FFurkan Gözükara SECourses if you get 7 GPU SXM machine with A100

wardensc2•10/16/24, 10:36 AM

So I can't use A6000 right ? because based your named 48gb I though 7xA6000 are enough

Carlos•10/16/24, 10:42 AM

Is Torch 2.5 OK now or not? I can't figure it out from the Video Tutorial, this is for Dreambooth Flux

Wwardensc2 So I can't use A6000 right ? because based your named 48gb I though 7xA6000 are ...

Furkan Gözükara SECourses•10/16/24, 12:28 PM

you can use

Furkan Gözükara SECourses•10/16/24, 12:28 PM

but not for fine tuning

Furkan Gözükara SECourses•10/16/24, 12:28 PM

but for lora multi gpu

CCarlos Is Torch 2.5 OK now or not? I can't figure it out from the Video Tutorial, this ...

Furkan Gözükara SECourses•10/16/24, 12:28 PM

it is ok to use for windows

Will Scarlett Ohara•10/16/24, 1:44 PM

https://1drv.ms/u/s!AoU1QUjSr6pJrbdRgnQNUMMyMe2qTg?e=xCOQuv
Hi, could you guide me through what's wrong with this training. My latest samples at 4400 steps are still far from imitating Justin Sweet's style. Thanks

WWill Scarlett Ohara https://1drv.ms/u/s!AoU1QUjSr6pJrbdRgnQNUMMyMe2qTg?e=xCOQuv Hi, could you guide ...

Furkan Gözükara SECourses•10/16/24, 2:34 PM

yep

Furkan Gözükara SECourses•10/16/24, 2:34 PM

crop all to 1024x1024

Furkan Gözükara SECourses•10/16/24, 2:34 PM

dont use captions

Furkan Gözükara SECourses•10/16/24, 2:35 PM

token can be like

Furkan Gözükara SECourses•10/16/24, 2:35 PM

ohwx painting style

FFurkan Gözükara SECourses crop all to 1024x1024

Will Scarlett Ohara•10/16/24, 2:38 PM

Would you have a way to automate croping/upscaling please, with ComfyUI for example?

WWill Scarlett Ohara Would you have a way to automate croping/upscaling please, with ComfyUI for exam...

Furkan Gözükara SECourses•10/16/24, 2:38 PM

i have auto cropper for some subjects

FFurkan Gözükara SECourses ohwx painting style

Will Scarlett Ohara•10/16/24, 2:39 PM

for the token it's here? like that?

Furkan Gözükara SECourses•10/16/24, 2:39 PM

https://www.patreon.com/posts/sota-subject-and-88391247

Patreon

SOTA Subject Cropper and Face Focused Image Resizer Scripts Do Bett...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

WWill Scarlett Ohara for the token it's here? like that?

Furkan Gözükara SECourses•10/16/24, 2:39 PM

yep

Will Scarlett Ohara•10/16/24, 2:40 PM

thank you

VigilenceOP•10/16/24, 4:37 PM

@Furkan Gözükara SECourses Hi Dr. Can you please share which video teaches me how to use huggingface for kohya_ss save state uploads/sync etc. ?

MrGTAmodsgerman•10/16/24, 9:27 PM

@Furkan Gözükara SECourses Bot attack again

daroth•10/16/24, 9:44 PM

hm, is Flux schnell trainable?

VVigilence <@205854764540362752> Hi Dr. Can you please share which video teaches me how to...

Furkan Gözükara SECourses•10/16/24, 9:45 PM

i couldnt make it work

Furkan Gözükara SECourses•10/16/24, 9:45 PM

i manually upload via notebook

Ddaroth hm, is Flux schnell trainable?

Furkan Gözükara SECourses•10/16/24, 9:45 PM

my results were bad

VigilenceOP•10/16/24, 10:19 PM

@Furkan Gözükara SECourses What is the correct folder path for the image? I would need to enter this in the resume save-state field.

VVigilence <@205854764540362752> What is the correct folder path for the image? I would ne...

Furkan Gözükara SECourses•10/16/24, 10:24 PM

i never made it work

VigilenceOP•10/16/24, 11:47 PM

I made it work in windows, but I don't know to get the proper folder path in ubuntu. In windows, if you select the two options in the picture, you can shutdown and resume based on the last save, with the same learning rate etc. Nothing is lost or needs to be done, which is preffered instead of just loading the last saved model file. You would then link to the folder of the laste save-state folder of the model you are training to resume. In onetrainer its much easier.

VVigilence I made it work in windows, but I don't know to get the proper folder path in ubu...

Furkan Gözükara SECourses•10/17/24, 8:56 AM

you can use last checkpoint

Furkan Gözükara SECourses•10/17/24, 8:56 AM

easier

Carlos•10/17/24, 9:02 AM

Quick question, does fine tuning require less steps than LORA?
My 3,000 steps fine tuning from last night (15 pics x 200 epochs) is giving much less good results than my 5,800 steps Dim 16 Lora I did with Fluxgym.
Is that normal?

wantedrbk•10/17/24, 10:02 AM

hi , just trained dreambooth 1 batch , runing swarmUI but on half proccess of generating img have this error , what it probably cause

error.txt3.62KB

wantedrbk•10/17/24, 10:14 AM

almost finished rendering but an error (

VigilenceOP•10/17/24, 11:06 AM

@Furkan Gözükara SECourses The use of save-states is much better as its actually continuing the training of the model. Starting off from the last saved checkpoint is essentially a whole different training of the model and the results will be different.

VigilenceOP•10/17/24, 11:09 AM

Anyway, I figured out how to use saved state and it worked, you just have to enable it at the start and it saves a special folder called saved-state. If for some reason you have to continue training due to crash or no HD space you can continue by inputting the folder patch into this field before you start. If you get an error then the path is not correct. If the folder exists with what is seen in the photo then it will definitely continue training with the same state it left off at as well as same result of training, just a continuation.

VigilenceOP•10/17/24, 11:26 AM

I forgot to mention, you would also have to edit the number of epochs, to be what you did not initially finish. So if you have a training for 200 epochs, and the saved state is from the 150 epoch, you have to change the epoch to 50 in the settings, or else it will continue training from the 150th epoch save-state for another 200 epochs. You can increase the epochs what ever you like, but it will be start from epoch 1 to what you set, since you are continuing the training using the saved-state folder. Its confusing if you don't understand this part.

<@205854764540362752> Have you looked into flux rapid lora training? Both fal and civit.ai offer i

Similar Threads