Software Engineering Courses (SECourses)•5mo ago

Hey guys, can confirm the tests of our dearest Dr Furkan. Flux Krea training is amazing but needs

Hey guys, can confirm the tests of our dearest Dr Furkan.

Flux Krea training is amazing but needs more epoch. I trained a makeup style checkpoints with 200+ images, and this is epoch 200. With flux dev the sweetspot is often around 140/150 epochs for my trainings.

I will train 20 epochs more on this one cause I think it may be beneficial for some details. Go ahead Flux Krea is awesome, reach for high epochs !

0950002-High_resolution_editorial_close-up_photo-beauty-200.png

Sakami•8/16/25, 10:38 AM

nice, whats your learning rate?

skyrrrOP•8/16/25, 10:40 AM

4e-06

Sskyrrr Hey guys, can confirm the tests of our dearest Dr Furkan. Flux Krea training i...

Furkan Gözükara SECourses•8/16/25, 10:47 AM

matching with my results

Furkan Gözükara SECourses•8/16/25, 10:48 AM

written this on patroen post :d

SSakami No LoRA vs LoRA Wan2.2

Sakami•8/16/25, 11:14 AM

@cyberbol

Sskyrrr Hey guys, can confirm the tests of our dearest Dr Furkan. Flux Krea training i...

K4bronsky•8/16/25, 12:25 PM

Glad to see this I’ll keep training

Sskyrrr Hey guys, can confirm the tests of our dearest Dr Furkan. Flux Krea training i...

K4bronsky•8/16/25, 12:47 PM

Have you trained a Lora on the same makeup style?

Joony•8/17/25, 1:49 AM

Hello Dr, do you think we will be able to do something similar to dreambooth style full finetune of qwen? Is this possible?

筋

筋斗云•8/17/25, 4:14 AM

I would like to ask everyone, I used 38 high-definition half-body portrait materials to fine-tune the DreamBooth training, with a configuration of 48G and 200 training rounds. Currently, the results from round 120 to round 200 are indeed becoming more similar to the characters in the material set, but I encountered two problems: 1. The test results for cartoon images are all realistic people, 2. If the character is relatively small in the image, there is only a very small probability that the face resembles the original.

JonkoXL•8/17/25, 7:44 AM

Did you caption the cartoon images with a cartoon keyword, and photorealism images with "photo" or a similar keyword?

JonkoXL•8/17/25, 7:46 AM

Maybe also introduce images where the face is more distant in your training ? Maybe you could pad the training images twice, but the second time with a big black border, so the face comes out relatively small?

JJoony Hello Dr, do you think we will be able to do something similar to dreambooth sty...

Furkan Gözükara SECourses•8/17/25, 8:08 AM

dreambooth later i think

Furkan Gözükara SECourses•8/17/25, 8:08 AM

first lora

筋筋斗云 I would like to ask everyone, I used 38 high-definition half-body portrait mater...

Furkan Gözükara SECourses•8/17/25, 8:09 AM

how do you test?=

Furkan Gözükara SECourses•8/17/25, 8:09 AM

2 : yes zoom in character

筋

筋斗云•8/17/25, 8:10 AM

comfyui，Grid test

筋

筋斗云•8/17/25, 8:11 AM

I relabeled the data again, added a realistic style, and plan to retrain for DB testing.

筋筋斗云 comfyui，Grid test

Furkan Gözükara SECourses•8/17/25, 8:12 AM

you are training a game character or something?

筋

筋斗云•8/17/25, 8:13 AM

I'm using the krea flux base model. I don't know if this is the reason.

筋

筋斗云•8/17/25, 8:14 AM

I'm training a LoRA of a character similar to you using my own photos. However, currently, when I write the cartoon prompt, it turns into a real - life person.

筋

筋斗云•8/17/25, 8:14 AM

According to your suggestion, I train in the form of DB to achieve the effect of LoRA.

筋筋斗云 I'm using the krea flux base model. I don't know if this is the reason.

Furkan Gözükara SECourses•8/17/25, 8:19 AM

oh you train real person and want cartoon output?

筋

筋斗云•8/17/25, 8:24 AM

I saw that there are cartoon - style characters in your videos.

筋筋斗云 I saw that there are cartoon - style characters in your videos.

Furkan Gözükara SECourses•8/17/25, 8:28 AM

yes they are hard

Furkan Gözükara SECourses•8/17/25, 8:29 AM

if you want cartoon flux kontext much better

Furkan Gözükara SECourses•8/17/25, 8:29 AM

https://youtu.be/adF9X9E0Chs

YouTubeSECourses

FLUX Kontext Dev Detailed Local Windows How To Tutorial - Better Th...

Black Forest Labs (BFL) publisher of FLUX models kept their promise and published the FLUX.1 Kontex DEV model today. So I didn't sleep and after doing huge research and test non-stop, I have prepared this excellent step by step tutorial that will show you how to use this amazing model. With FLUX Kontext you can edit any part of the image in any ...

Furkan Gözükara SECourses•8/17/25, 8:30 AM

and flux krea will make it even harder

Furkan Gözükara SECourses•8/17/25, 8:30 AM

for cartoon

筋

筋斗云•8/17/25, 8:31 AM

I can use kontext, and I have successfully trained my own control lora. At present, I just want to replicate the process in your tutorial, use a dataset of real - person photos to train a DreamBooth (DB) to achieve a personal model with generalization ability.

筋

筋斗云•8/17/25, 8:32 AM

Thank you for your reply.

筋筋斗云 I can use kontext, and I have successfully trained my own control lora. At prese...

Furkan Gözükara SECourses•8/17/25, 9:11 AM

use flux dev not krea for this task

Furkan Gözükara SECourses•8/17/25, 9:11 AM

flux krea is optimized for realism

筋

筋斗云•8/17/25, 9:31 AM

Thank you for your suggestion. I'm re - modifying the main model for multi - GPU retraining.

筋筋斗云 Thank you for your suggestion. I'm re - modifying the main model for multi - GPU...

Furkan Gözükara SECourses•8/17/25, 9:32 AM

for multi gpu loras works great with 48 gb gpus

Furkan Gözükara SECourses•8/17/25, 9:32 AM

dreambooth need 80 gb gpus

筋

筋斗云•8/17/25, 11:14 AM

4X 4090 48G

筋筋斗云 4X 4090 48G

Furkan Gözükara SECourses•8/17/25, 7:47 PM

awesome

Furkan Gözükara SECourses•8/17/25, 7:47 PM

i wish they were sold worldwide

Furkan Gözükara SECourses•8/17/25, 7:47 PM

definitely better than 32 gb 5090

Timson•8/17/25, 10:44 PM

Tried training character checkpoint with 50 image set 7000 steps from flux krea and result is much worse than flux dev counterpart so far. Much less likeness, seems more weird results

TTimson Tried training character checkpoint with 50 image set 7000 steps from flux krea ...

Furkan Gözükara SECourses•8/18/25, 12:31 AM

i feel krea need more epochs

TTimson Tried training character checkpoint with 50 image set 7000 steps from flux krea ...

K4bronsky•8/18/25, 1:00 AM

I found using a character Lora that I trained with flux on flux krea is giving me the textures of krea with the character. Vs the finetune I was getting fuzzy images

Tim0807•8/18/25, 1:05 AM

is a 80gb gpu needed for flux krea Dreambooth/Finetune?

As far as I remember for flux dev DB/Finetune a 48gb gpu was enough

TTim0807 is a 80gb gpu needed for flux krea Dreambooth/Finetune? As far as I remember f...

Furkan Gözükara SECourses•8/18/25, 1:18 AM

6 gb sufficient

Furkan Gözükara SECourses•8/18/25, 1:18 AM

if you want multiple gpu yes

zadkhamaya•8/18/25, 7:54 AM

Hi,

I was trying to generate using my trained dreambooth model on Runpod. It shown this error in SwarmUI. When I run it in massedcompute, it's working fine. Do you what is the cause of this error?

Thank you

Zzadkhamaya Hi, I was trying to generate using my trained dreambooth model on Runpod. It sh...

Furkan Gözükara SECourses•8/18/25, 11:49 AM

model download failed

Furkan Gözükara SECourses•8/18/25, 11:49 AM

rm -r modelpath

Furkan Gözükara SECourses•8/18/25, 11:49 AM

redownload

skyrrrOP•8/18/25, 1:58 PM

@Dr. Furkan Gözükara Hey Dr, is there a theoratical limit with checkpoint training regarding dataset ? I have a good 350 images, is that too much or I would just need to train it longer ?

Hey guys, can confirm the tests of our dearest Dr Furkan. Flux Krea training is amazing but needs

Similar Threads

Similar Threads

Similar Threads