Software Engineering Courses (SECourses)•5mo ago

so whats better for fine tuned image gen realism right now? dev or krea? pros and cons? can someo

so whats better for fine tuned image gen realism right now?

dev or krea?

pros and cons? can someone give me their observations?

JJamiewhelton so whats better for fine tuned image gen realism right now? dev or krea? pros...

Furkan Gözükara SECourses•8/16/25, 8:13 AM

i posted grids and hopefully will make a video today

Furkan Gözükara SECourses•8/16/25, 8:13 AM

still working on wan 2.2 presets

JamiewheltonOP•8/16/25, 8:17 AM

where can I see your grids, patreon?

JamiewheltonOP•8/16/25, 8:18 AM

and how about wan 2.2? kling 2.1 master still better?

JJamiewhelton and how about wan 2.2? kling 2.1 master still better?

Furkan Gözükara SECourses•8/16/25, 8:18 AM

i dont use kling sadly

JJamiewhelton and how about wan 2.2? kling 2.1 master still better?

Furkan Gözükara SECourses•8/16/25, 8:18 AM

girds here https://www.patreon.com/posts/112099700

Patreon

Kohya FLUX Fine Tuning (Full Checkpoints) Training Full Tutorial Fo...

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

Furkan Gözükara SECourses•8/16/25, 8:19 AM

JamiewheltonOP•8/16/25, 8:22 AM

I appreciate you taking the time to make that

JamiewheltonOP•8/16/25, 8:22 AM

I personally don't see any differences/patterns

JamiewheltonOP•8/16/25, 8:22 AM

probably going to stick to flux.dev for now

skyrrr•8/16/25, 10:35 AM

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?

skyrrr•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...