Software Engineering Courses (SECourses)•10mo ago

Hi! This guy can apparently train dreambooth models on 12 gb vram, any thoughts on how that might be

Hi! This guy can apparently train dreambooth models on 12 gb vram, any thoughts on how that might be possible? https://www.linkedin.com/pulse/comprehensive-guide-midjourney-sref-style-training-4-valsecchi-cxi6f/?trackingId=ihrUD9hrgJh%2F1CaCHPBg%2BA%3D%3D

spt999•3/3/25, 4:59 PM

@Furkan Gözükara SECourses i finished learning with 2 batch size, it is good speed or not? Thank you for anser and help

message.txt29.48KB

EElli_Oaken Hi! This guy can apparently train dreambooth models on 12 gb vram, any thoughts ...

Furkan Gözükara SECourses•3/3/25, 6:47 PM

we can dreambooth on 6 gb vram

Furkan Gözükara SECourses•3/3/25, 6:47 PM

https://youtu.be/FvpWy1x5etM

YouTubeSECourses

FLUX Full Fine-Tuning / DreamBooth Training Master Tutorial for Win...

If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...

Sspt999 <@205854764540362752> i finished learning with 2 batch size, it is good speed o...

Furkan Gözükara SECourses•3/3/25, 6:47 PM

great speed

Furkan Gözükara SECourses•3/3/25, 6:47 PM

because batch size 2 = 2x speed

FFurkan Gözükara SECourses https://youtu.be/FvpWy1x5etM

Elli_OakenOP•3/3/25, 6:51 PM

So confused, thought you needed 80 vram

EElli_Oaken So confused, thought you needed 80 vram

Furkan Gözükara SECourses•3/3/25, 6:51 PM

it is very old info

Furkan Gözükara SECourses•3/3/25, 6:51 PM

now block swap technique is well developed and established

Furkan Gözükara SECourses•3/3/25, 6:51 PM

all you need is 64 GB RAM + 6 GB VRAM

FFurkan Gözükara SECourses all you need is 64 GB RAM + 6 GB VRAM

Elli_OakenOP•3/3/25, 6:53 PM

Woops okey, does it work on kohya v46?

EElli_Oaken Woops okey, does it work on kohya v46?

Furkan Gözükara SECourses•3/3/25, 6:54 PM

yes

Furkan Gözükara SECourses•3/3/25, 6:55 PM

latest zip file on patreon are up to date

FFurkan Gözükara SECourses great speed

spt999•3/3/25, 6:59 PM

Thank you! Results of generation are amazing even without yolo for now, thank you very much for guide, tomorrow will try with yolo. This is truly the best guide

Sspt999 Thank you! Results of generation are amazing even without yolo for now, thank yo...

Furkan Gözükara SECourses•3/3/25, 7:02 PM

Thanks a lot. I recently researched a better workflow but this performed best

Sspt999 Thank you! Results of generation are amazing even without yolo for now, thank yo...

Nipunn•3/4/25, 3:51 PM

Sounds good

FFurkan Gözükara SECourses Thanks a lot. I recently researched a better workflow but this performed best 🙂

spt999•3/4/25, 6:30 PM

With yolo incredible results very good! I jusst don't understand how you gain approximately spee 1.3 s/it on 3090, i saw post on X.com

Sspt999 With yolo incredible results very good! I jusst don't understand how you gain ap...

Furkan Gözükara SECourses•3/4/25, 6:34 PM

i can implement flash attention - may help. also teacache

Furkan Gözükara SECourses•3/4/25, 6:34 PM

but both may reduce quality

Furkan Gözükara SECourses•3/4/25, 6:35 PM

but it is already that speed on my rtx 3090 ti

Furkan Gözükara SECourses•3/4/25, 6:35 PM

FFurkan Gözükara SECourses i can implement flash attention - may help. also teacache

spt999•3/4/25, 7:19 PM

thank you i will try tomorrow

spirit18•3/5/25, 8:44 PM

I see movement in GPU and RAM usage

spirit18•3/5/25, 8:45 PM

Sspirit18 I see movement in GPU and RAM usage

Furkan Gözükara SECourses•3/5/25, 8:48 PM

which config?

Sspirit18 I see movement in GPU and RAM usage

Furkan Gözükara SECourses•3/5/25, 8:49 PM

your config wrong

Furkan Gözükara SECourses•3/5/25, 8:49 PM

8-bit AdamW optimizer

Furkan Gözükara SECourses•3/5/25, 8:49 PM

please follow tutorial dont skipo

Furkan Gözükara SECourses•3/5/25, 8:50 PM

https://youtu.be/FvpWy1x5etM

YouTubeSECourses

FLUX Full Fine-Tuning / DreamBooth Training Master Tutorial for Win...

spirit18•3/5/25, 8:55 PM

Okay will ReDo

spt999•3/7/25, 2:26 PM

Hello @Furkan Gözükara SECourses , after all tries i have few questions:

if i use 15 batch size on a100 i need use approximately reduce x2 epochs? I have 137 photo, i think need near 100 epochs with 7 batch size, if batch size 15 can try approx 60 epochs?
if my photos from iphone them size approx 1 mb, yours photo are 5mb is it very important in this case, i dont see very big difference in learning (using 137 preprocessed photos) avr_loss near 0.4 always +-
Now i'm doing upscaling with SUPIR, and will try after it will finish.
tell please can i fix freezes with scrolling screen with ThinLink macos m4 max?

spirit18•3/7/25, 5:38 PM

Huge Thanks to @Furkan Gözükara SECourses

2230-a_propoganda_style_poster_from_nazi_germ-model37_abygen1-1220266658.png

2220-photograph_of_ohwx_man_sitting_in_front-model37_abygen1-000100-1707356540.png

2211-photograph_of_ohwx_man_riding_a_majestic-model37_abygen1-000100-1495114612.png

Sspt999 Hello <@205854764540362752> , after all tries i have few questions: - if i use ...

Furkan Gözükara SECourses•3/7/25, 10:19 PM

epoch doesnt change according to batch size

Furkan Gözükara SECourses•3/7/25, 10:20 PM

epoch is related to number of image you have usually

Furkan Gözükara SECourses•3/7/25, 10:20 PM

mb of photos doesnt matter what matter is resolution like 1024x1024 - best

Furkan Gözükara SECourses•3/7/25, 10:20 PM

if you dont downscale it will auto

Furkan Gözükara SECourses•3/7/25, 10:20 PM

sadly thinlinc client is not very good. @NicB @computersheep are there any better alternative?

Sspirit18 Huge Thanks to <@205854764540362752> 🥹

Furkan Gözükara SECourses•3/7/25, 10:20 PM

awesome

computersheep•3/7/25, 10:23 PM

There are many alternatives just not ones we can preinstall for our users.
The few free alternatives I tried myself felt worse.
There are some paid vnc clients that are pretty good but again none we can provide, users would have to signup and install those themselves.
I suspect all clients will be mediocre for users outside of the USA due to latency though.

computersheep•3/7/25, 10:24 PM

If you just want better file syncing, I personally use winscp on windows, works pretty well and much faster.
FileZilla is also pretty good, faster than winscp, but I like the winscp ui better for simple folder syncs.

Ccomputersheep If you just want better file syncing, I personally use winscp on windows, works ...

Furkan Gözükara SECourses•3/7/25, 10:27 PM

i think people rather looking for more fluent working connection interface

computersheep•3/7/25, 10:28 PM

Thinlinc is the best for that. All other solutions require more work for users to install and setup.

computersheep•3/7/25, 10:29 PM

Here is example of a product that would be more performant but users would need to install and setup themselves.

https://www.realvnc.com/

RealVNC®Jan Intia

RealVNC® - Remote access software for desktop and mobile | RealVNC

Secure, cost-effective remote access and support for your computer, from your desktop or mobile device. VNC® Connect enables cloud or direct connectivity.

Ccomputersheep Thinlinc is the best for that. All other solutions require more work for users t...

Furkan Gözükara SECourses•3/7/25, 10:30 PM

i see thanks

Pixel.dust•3/8/25, 3:21 AM

I just wanted to share what I did, using the tips and feedback I received from the people here, can I ask if there is a proper method of evaluate models

https://www.linkedin.com/posts/italo-thiago-3b3158216_illustriousxl-deeplearning-gpucomputing-activity-7303857523596414976-fgrY?utm_source=social_share_send&utm_medium=android_app&rcm=ACoAADZ7uTcBFffVIJxesBcJt71j4tV36U6H7pk&utm_campaign=copy_link

#illustriousxl #deeplearning #gpucomputing #stablediffusion… | Ital...

Just finished a preliminary training run with Illustrious XL using localy 2xRTX 3090s. 36 hours, 2 epochs, and 1344 CC0 images, Here some exciting initial results. The speed and quality are impressive! More in-depth analysis to follow. #IllustriousXL #DeepLearning #GPUComputing #StableDiffusion #DomesticCluster #AIart #GrapichDesign #Genai