Software Engineering Courses (SECourses)•11mo ago

@Dr. Furkan Gözükara which model is better for fine tuning with only 130 pics of a single person ad

@Dr. Furkan Gözükara which model is better for fine tuning with only 130 pics of a single person adn the dataset doesn't have much variations . Is SDXL+Dreambooth a good combo or FluxDev+Drreambooth a good combo.I want the best realistic photos

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Can you test prodigy as well?

Furkan Gözükara SECourses•2/17/25, 9:15 AM

i tried but doesnt fit into 48 gb

Pphoenix @Dr. Furkan Gözükara which model is better for fine tuning with only 130 pics o...

Furkan Gözükara SECourses•2/17/25, 9:15 AM

flux is always the best atm

FFurkan Gözükara SECourses i tried but doesnt fit into 48 gb

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/17/25, 9:15 AM

for dreambooth or lora?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 for dreambooth or lora?

Furkan Gözükara SECourses•2/17/25, 9:16 AM

dreambooth

FFurkan Gözükara SECourses dreambooth

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/17/25, 9:16 AM

Are you not testing lora?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Are you not testing lora?

Furkan Gözükara SECourses•2/17/25, 9:17 AM

nope

phoenixOP•2/17/25, 9:34 AM

@Dr. Furkan Gözükara so I was able to get 160 High Quality edited and retouched pics of the same person and 100 RAW ones directly from the camera .

Will it be a good idea to mix both types of pics as the edited ones have only face as the focus but the raw ones do have different body angles

So will it be better if I train Flux+Dreambooth with 160 edited+100 raw pics.total dataset of 260 pics

Pphoenix @Dr. Furkan Gözükara so I was able to get 160 High Quality edited and retouched ...

Furkan Gözükara SECourses•2/17/25, 9:37 AM

it depends on purpose

Furkan Gözükara SECourses•2/17/25, 9:37 AM

if you want to get both outputs after training

Furkan Gözükara SECourses•2/17/25, 9:37 AM

yes you can mix

Furkan Gözükara SECourses•2/17/25, 9:46 AM

paged optimizers are a great feature

Furkan Gözükara SECourses•2/17/25, 9:46 AM

they utilize VRAM to maximum

Furkan Gözükara SECourses•2/17/25, 9:46 AM

those used ones are running training atm

phoenixOP•2/17/25, 10:01 AM

@Dr. Furkan Gözükara I mean can the fine tuning give good output,I want only single output I want retouched images only after outpit so if the output is not accurate sometimes,I can just use SUPR to upscale

Is this the right approach?

phoenixOP•2/17/25, 10:03 AM

Can I use this flux model for training or the you told about in the video is the right one?

Screenshot_2025-02-17-15-33-15-338_org.mozilla.firefox.jpg

chomsky honkerton•2/17/25, 10:06 AM

im watching a video for sdxl lora training rn where my boy says everything is for free, but the SDXL configs are behind a pay wall. can i find them here?
https://www.youtube.com/watch?v=0t5l6CP9eBg

YouTubeSECourses

Full Stable Diffusion SD & XL Fine Tuning Tutorial With OneTrainer ...

In this tutorial, I am going to show you how to install OneTrainer from scratch on your computer and do a Stable Diffusion SDXL (Full Fine-Tuning 10.3 GB VRAM) and SD 1.5 (Full Fine-Tuning 7GB VRAM) based models training on your computer and also do the same training on a very cheap cloud machine from MassedCompute if you don't have such compute...

Cchomsky honkerton im watching a video for sdxl lora training rn where my boy says everything is fo...

Furkan Gözükara SECourses•2/17/25, 10:06 AM

the video displays them keep watching

FFurkan Gözükara SECourses the video displays them keep watching

chomsky honkerton•2/17/25, 10:06 AM

ok, thank you so much - will do

Pphoenix Can I use this flux model for training or the you told about in the video is the...

Furkan Gözükara SECourses•2/17/25, 10:06 AM

yes this model best

Pphoenix @Dr. Furkan Gözükara I mean can the fine tuning give good output,I want only sin...

Furkan Gözükara SECourses•2/17/25, 10:07 AM

in that case dont include retouched images

chomsky honkerton•2/17/25, 10:07 AM

i just bought a 3090, trying it out for the first time with your tutorials

FFurkan Gözükara SECourses yes this model best

phoenixOP•2/17/25, 10:11 AM

@Dr. Furkan Gözükara so the models command in the terminal automatically fetches this one right?

Pphoenix @Dr. Furkan Gözükara so the models command in the terminal automatically fetches...

Furkan Gözükara SECourses•2/17/25, 10:13 AM

if you are using our downloaders yes

Furkan Gözükara SECourses•2/17/25, 10:13 AM

flux 1 dev

phoenixOP•2/17/25, 10:13 AM

Ok thanks

Furkan Gözükara SECourses•2/17/25, 10:39 AM

i hate slowness of runpod

Furkan Gözükara SECourses•2/17/25, 10:40 AM

you waste at least 5x time to load and install something compared to massed compute

Furkan Gözükara SECourses•2/17/25, 10:40 AM

sometimes even more

FFurkan Gözükara SECourses you waste at least 5x time to load and install something compared to massed com...

phoenixOP•2/17/25, 10:43 AM

Yeah massedcompute is really fast,I like it

phoenixOP•2/17/25, 11:00 AM

@Dr. Furkan Gözükara i want my model to get the body structure also correct so can I put some images in the dataset without cropping ,only resizing

Cchomsky honkerton im watching a video for sdxl lora training rn where my boy says everything is fo...

chomsky honkerton•2/17/25, 11:17 AM

if i am trying to create a character for myself like was done in these video, is it best to set this drop down in the top right to LoRa?

chomsky honkerton•2/17/25, 11:17 AM

in the video it was on fine tune

chomsky honkerton•2/17/25, 11:18 AM

and would that effect what settings i would want to use

Cchomsky honkerton if i am trying to create a character for myself like was done in these video, is...

chomsky honkerton•2/17/25, 11:32 AM

I’m also curious about number of images. I have 59 in my dataset. I noticed in the video you did 15 images x 200 epochs = 3000 steps.

I went ahead and did 59 x 100 epochs = 5900 steps and will look extra carefully at the models around epoch 5, I kinda just wanted the slack of the other epochs in case it’s under trained by then. thoughts on this? how many epochs would you go with? seems like with my hardware I’ll hit epoch 100/100 in 4-5 hrs

chomsky honkerton•2/17/25, 11:33 AM

def got 9 bucks worth of info out of the video and I’m not even done yet lol, cheers mate

Pphoenix Yeah massedcompute is really fast,I like it

Furkan Gözükara SECourses•2/17/25, 12:02 PM

yep

Pphoenix @Dr. Furkan Gözükara i want my model to get the body structure also correct so c...

Furkan Gözükara SECourses•2/17/25, 12:02 PM

yes have more images for that

Cchomsky honkerton if i am trying to create a character for myself like was done in these video, is...

Furkan Gözükara SECourses•2/17/25, 12:02 PM

i dont recommend lora

Furkan Gözükara SECourses•2/17/25, 12:02 PM

dreambooth better

Furkan Gözükara SECourses•2/17/25, 12:02 PM

then you can extract lora

phoenixOP•2/17/25, 12:10 PM

@Dr. Furkan Gözükara I have messaged you please check

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/17/25, 12:35 PM

@Dr. Furkan Gözükara Is there a way to save epochs mid training? Like I have set up save evry 25 epoch but while training I want to save the 10th epoch as well?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 @Dr. Furkan Gözükara Is there a way to save epochs mid training? Like I have set...

Furkan Gözükara SECourses•2/17/25, 12:44 PM

i dont know

Pphoenix @Dr. Furkan Gözükara I have messaged you please check

Furkan Gözükara SECourses•2/17/25, 12:44 PM

checking