Software Engineering Courses (SECourses)•11mo ago

Maybe that was the difference then as I trained the LoRAs with AdamW but the dreambooth finetune wit

Maybe that was the difference then as I trained the LoRAs with AdamW but the dreambooth finetune with Adafactor (I could not fit AdamW in my gpu or a rented 48GB one)
Do you think this could be the difference?
I still believe finetuning should be better so I’m just confused about why LoRA yields better results for me.

DDario Maybe that was the difference then as I trained the LoRAs with AdamW but the dre...

Furkan Gözükara SECourses•2/11/25, 9:26 PM

hard to know possibly a mistake somewhere

Aleksei Naumov•2/11/25, 10:05 PM

Hey there!
Anyone out here was deploying flux lora training as an API?
Would be happy to chat or be grateful if you can share something useful on how to deploy to runpod/other compute providers.
P.S. If you can share some useful presets for training person lora - beer on me haha!

Hirmuolio•2/11/25, 11:31 PM

Does training lora for Noobai v-pred require different settings?
I'm using kohya_ss gui (dev branch that is up to date with the latest script).
Someone somewhere said that it needs

--zero_terminal_snr

--zero_terminal_snr

parameter. Anything else?

HHirmuolio Does training lora for Noobai v-pred require different settings? I'm using kohya...

Furkan Gözükara SECourses•2/12/25, 12:20 AM

what is Noobai v-pred

FFurkan Gözükara SECourses i dont see any reason to use it for individuals

wardensc2•2/12/25, 3:40 AM

recently when people train lora with kohya and they also train clip text encode together, they claim that lora train together with clip text encode give much better result. I wonder your latest lora training script json file have these clip text encode trained with lora or not ?

FFurkan Gözükara SECourses what is Noobai v-pred

wardensc2•2/12/25, 3:42 AM

It a finetuning of Illustrious model, another heavy modified training from SDXL model.

Hirmuolio•2/12/25, 5:09 AM

NoobAI-XL v-pred uses v-pred instead of eps like other SDXL models which makes it behave a bit different.

FFurkan Gözükara SECourses hard to know possibly a mistake somewhere 😄

DarioOP•2/12/25, 9:36 AM

I'll keep testing and let you know. In general the deambooth finetune does preserve small details betetr (logos, brake caliper, etc), but the LoRA follows the car's shape a lot better. I need to test it further with a contolnet and depth map (taken from the 3D model) to see if this car body issue can be resolved that way. And maybe as a final test I can send a 80GB+ GPU to do a AdamW dreambooth training run.

Hirmuolio•2/12/25, 9:49 AM

After a test I can say that normal SDXL lora training settings +

--zero_terminal_snr

--zero_terminal_snr

on NoobAI v-pred doesn't work. Lora outputs just grey images that sometimes have very blurry subject.

DDario Hi Furkan, I've been experimenting with Flux dreambooth and LoRA training for se...

Timson•2/12/25, 9:50 AM

are you sure fine tune checkpoint is not just undertrained? How many images do you have in the trainign dataset and how many training steps did you do for lora and fine-tuning?

Timson•2/12/25, 9:51 AM

The thing is - lora training is just much faster in terms of steps/resemblance compared to full checkpoint, but checkpoint should yeild better generalization

AAleksei Naumov Hey there! Anyone out here was deploying flux lora training as an API? Would be ...

Timson•2/12/25, 9:53 AM

this chat is full of friendly people sharing their knowledge of training flux, for starters just follow basic tutorials of Dr. Furkan on the subject you are interested, they cover 99% of your questions

TTimson are you sure fine tune checkpoint is not just undertrained? How many images do y...

DarioOP•2/12/25, 10:12 AM

Good point, thanks for your suggestion. The dataset has 45 images. LoRA produces good results at around 6000 steps. For dreambooth I am trying around 9 - 10K steps. Do you think I should train longer still?

DDario Good point, thanks for your suggestion. The dataset has 45 images. LoRA produces...

Timson•2/12/25, 10:18 AM

yes, continue training if you don't see model sanity degradation, you can get better results compared to lora at 18-20k steps or more, if your use case require perfect resemblance at the cost of lower model sanity

Timson•2/12/25, 10:19 AM

but if your use case do not benefit much from improved generalization this can be just too costly for similar results

TTimson yes, continue training if you don't see model sanity degradation, you can get be...

DarioOP•2/12/25, 10:25 AM

Great thanks I will give that a try. Costs and speed are not a priority, obtaining the best product accuracy / quality is, so I will give this a go by training a lot longer and compare the epochs in a grid. Cheers

Wwardensc2 recently when people train lora with kohya and they also train clip text encode ...

Furkan Gözükara SECourses•2/12/25, 11:44 AM

we already train clip

Furkan Gözükara SECourses•2/12/25, 11:44 AM

our all lora configs have it

DarioOP•2/12/25, 2:37 PM

Would it help to also train the text encoder (at 50% unet LR for example) and use T5 attention mask for dreambooth training just as we do with LoRA training? Has anyone seen any benefits?

FFurkan Gözükara SECourses we already train clip

wardensc2•2/12/25, 2:51 PM

Compare with train checkpoint follow your config, if I extract lora from trained checkpoint, is the extracted lora quality better ?

gio3000.•2/12/25, 3:10 PM

Annybody finetuned a style using image data set of a particular city district? I will start It over a previous finetuned person using dreambooth

DDario Would it help to also train the text encoder (at 50% unet LR for example) and us...

Furkan Gözükara SECourses•2/12/25, 3:28 PM

kohya not supporting

FFurkan Gözükara SECourses kohya not supporting

DarioOP•2/12/25, 3:47 PM

What do you mean? Both parameters can be applied in the koyha UI (dreambooth tab) and the training runs without issues. Do you mean it will cause problems with the final trained models?

DDario What do you mean? Both parameters can be applied in the koyha UI (dreambooth tab...

Furkan Gözükara SECourses•2/12/25, 9:46 PM

kohya doesnt train clip large or t5 when doing fine tuning

Furkan Gözükara SECourses•2/12/25, 9:46 PM

try you will see same speed and vram

FFurkan Gözükara SECourses try you will see same speed and vram

DarioOP•2/13/25, 9:04 AM

Ah I see, thanks for clarifying

DDario Ah I see, thanks for clarifying

Furkan Gözükara SECourses•2/13/25, 11:00 AM

yes i told to add kohya but he still didnt add

FFurkan Gözükara SECourses yes i told to add kohya but he still didnt add 😦

DarioOP•2/13/25, 1:42 PM

Interesting to test when it does get implemented at some point.

DDario Interesting to test when it does get implemented at some point.

Furkan Gözükara SECourses•2/13/25, 2:19 PM

100%

keyser_soze•2/14/25, 5:38 AM

@Furkan Gözükara SECourses Hi sir, we trained the Lora model using a quality Tier 1 config file with 48GB of VRAM.
We generated an AI character with photo grid, cropped, flipped, and upscaled the images. We created a dataset of about 8 images to train the Lora model.
While generating images from Lora, the close-ups have a really good face, but in wide shots, the face becomes pixelated.
Additionally, the face does not closely resemble the training dataset.
I’m attaching the images as well.
brown background image is generated
white background image is dataset

WhatsApp_Image_2025-02-14_at_11.01.31_AM.jpeg

WhatsApp_Image_2025-02-14_at_11.01.14_AM_1.jpeg

WhatsApp_Image_2025-02-14_at_11.01.14_AM.jpeg

Kkeyser_soze <@205854764540362752> Hi sir, we trained the Lora model using a quality Tier 1 ...

TheToday•2/15/25, 12:07 AM

I think the dataset is not good, the same background and clothes in all the photos? Probably the problem is in the dataset

TheToday•2/15/25, 12:08 AM

But I don't know, sometimes these things happen

TheToday•2/15/25, 12:10 AM

How many photos did you train?

TTheToday I think the dataset is not good, the same background and clothes in all the phot...

keyser_soze•2/15/25, 2:26 AM

8 images , actually

keyser_soze•2/15/25, 2:34 AM

Which is the best way to train a AI generated character then

keyser_soze•2/15/25, 2:37 AM

And which is the best way to create a dataset for AI generated character?, I tried multiple things, the consistent is not so good

keyser_soze•2/15/25, 2:57 AM

Can someone help me in this it will be really helpfull

Kkeyser_soze Can someone help me in this it will be really helpfull

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/15/25, 10:51 AM

try creating consistent images with pulid first and then train on them

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 try creating consistent images with pulid first and then train on them

keyser_soze•2/15/25, 10:55 AM

thanks nitin , for your suggestion , will try it ,and show the results here

Kkeyser_soze Which is the best way to train a AI generated character then

Furkan Gözükara SECourses•2/15/25, 11:33 AM

i recommend 2 stage

Furkan Gözükara SECourses•2/15/25, 11:33 AM

first train

Furkan Gözükara SECourses•2/15/25, 11:33 AM

generate better images

Furkan Gözükara SECourses•2/15/25, 11:33 AM

then collect more images

Furkan Gözükara SECourses•2/15/25, 11:33 AM

and train again

FFurkan Gözükara SECourses and train again

keyser_soze•2/15/25, 11:37 AM

Should I train the first trained Lora with a new dataset using more steps, or should I train from scratch using the images generated by the trained Lora as a dataset?

Kkeyser_soze Should I train the first trained Lora with a new dataset using more steps, or sh...

Furkan Gözükara SECourses•2/15/25, 11:49 AM

from scractch with new dataset

FFurkan Gözükara SECourses from scractch with new dataset

keyser_soze•2/15/25, 12:03 PM

cool

MetaKan•2/15/25, 12:55 PM

hi, getting this error installing Kohya_FLUX_DreamBooth_v17. Anyone can help?

Unc0nn3cteD•2/15/25, 2:04 PM

Hello Furkan, everyone!

I'm trying to get my head around the structure of the Patreon as well as tutorials and have a bit of a hard time to see where to start - will appreciate a pointer to a 101 or a tutorial

My use case WAS => Stable Diffusion 1.5 and some public CivitAI models + my own photography dataset with poses, light and bokehs (150-200 images) = New checkpoint + Dreambooth fine-tuning on a new person (50-100 images) = New checkpoint, that let me create portraits of people in my own recognizable photography style (Rembrandt light, bokehs, ethereal feeling). I used https://github.com/TheLastBen/fast-stable-diffusion workbook on Colab for both trainings (SD15/CIVIT + MyStyle-XYZ.jpg x 200 images) + (Person-XYZ x 100 images). Sample results are attached - these look exactly as real people to the extent the parents can't say if it is made with AI or not. With some experimenting I managed to get to 768 and 896 pixel resolutions here, all with Automatic1111.

What I want is to elevate this to SDXL with 1280 pixels and test out SDXL and CivitAI checkpoints. So, I would have 150-200 1024-1280px images to train my style and then 50-100 images of a person to generate images like (beautiful personXYZ as styleXYZ .... ). I can't find a reliable workbook (Colab or RunPod, paid are OK) with clear instructions. I have a good development background and PhD in CS, but in a slightly different area, so I struggle to get to the point where I have a code that works so that I could experiment. Like I found configs, but how do I set up the environment / deployed OneTrainer - but it won't accept .safetensors model / managed to adjust fast-stable-diffusion to accept SDXL, but it fails to work with GDrive

Any clues, hints or pointers I could use? Thank you so much in advance!

Cheers, Alex

MMetaKan hi, getting this error installing Kohya_FLUX_DreamBooth_v17. Anyone can help?

Furkan Gözükara SECourses•2/15/25, 2:21 PM

yes

Maybe that was the difference then as I trained the LoRAs with AdamW but the dreambooth finetune wit

Similar Threads