Software Engineering Courses (SECourses)•5mo ago

Man thank you so much I got confused by the many negative feedbacks regarding krea training. From yo

Man thank you so much I got confused by the many negative feedbacks regarding krea training. From your tests it seems to be a huge improvement over flux.dev. Do you think it would do well with larger datasets as well (200 images ?)

AiInfluence•8/13/25, 8:57 AM

there is

AiInfluence•8/13/25, 8:57 AM

dont listen to what they say about krea

Sskyrrr Man thank you so much I got confused by the many negative feedbacks regarding kr...

Furkan Gözükara SECourses•8/13/25, 8:57 AM

well it should i think. i trained flux dev with 256 images

FFurkan Gözükara SECourses at some prompts it works better than flux dev at some prompts worse

Sakami•8/13/25, 9:04 AM

I feel like its harder with makin more dramatic lighting

SSakami I feel like its harder with makin more dramatic lighting

Furkan Gözükara SECourses•8/13/25, 9:05 AM

yes it is biased to certain realism

skyrrrOP•8/13/25, 9:11 AM

Your training at LR 2e-6 is also quite good, on a fence between this one and the base 4e-06

Sskyrrr Your training at LR 2e-6 is also quite good, on a fence between this one and the...

Furkan Gözükara SECourses•8/13/25, 9:13 AM

yep

Furkan Gözükara SECourses•8/13/25, 9:19 AM

this kl_optimal is also nice scheduler

Furkan Gözükara SECourses•8/13/25, 9:20 AM

Furkan Gözükara SECourses•8/13/25, 9:21 AM

i will post grids

FFurkan Gözükara SECourses this kl_optimal is also nice scheduler

skyrrrOP•8/13/25, 9:37 AM

kl-optimal is my absolute favorite, especially in combinaison with gradient estimation

Sskyrrr kl-optimal is my absolute favorite, especially in combinaison with gradient esti...

Furkan Gözükara SECourses•8/13/25, 9:50 AM

which sampler it works best with? with IPNDM not very good

Furkan Gözükara SECourses•8/13/25, 9:50 AM

i guess i have to test

Furkan Gözükara SECourses•8/13/25, 9:52 AM

testing right now

FFurkan Gözükara SECourses also will publish results of flux dev vs flux krea dev

JonkoXL•8/13/25, 9:55 AM

Awesome!
Will you publish the lora or dataset you made of yourself too?

JJonkoXL Awesome! Will you publish the lora or dataset you made of yourself too? 😛

Furkan Gözükara SECourses•8/13/25, 9:57 AM

nope

Furkan Gözükara SECourses•8/13/25, 9:57 AM

also published comparison results on patreon post for dreambooth

JonkoXL•8/13/25, 9:59 AM

I'll just train on this guy then

https://cdn.discordapp.com/attachments/1042802773365891173/1404539825293361272/e487b900_merged_2images_top_bottom.jpg?ex=689d8965&is=689c37e5&hm=5edeb512da507288794c38bedfa2b30dacdb82eb98e17849f726a9ed7c5158f9&

JJonkoXL I'll just train on this guy then 🤣 https://cdn.discordapp.com/attachments/1042...

Furkan Gözükara SECourses•8/13/25, 10:00 AM

cyberbol•8/13/25, 10:20 AM

Very funny.

FFurkan Gözükara SECourses which sampler it works best with? with IPNDM not very good

skyrrrOP•8/13/25, 10:23 AM

It works best with uni_pc and gradient estimation

Sskyrrr It works best with uni_pc and gradient estimation

Furkan Gözükara SECourses•8/13/25, 10:31 AM

i am making a very big comparison

Furkan Gözükara SECourses•8/13/25, 10:32 AM

Furkan Gözükara SECourses•8/13/25, 10:32 AM

and comparing with our best

JJonkoXL I'll just train on this guy then 🤣 https://cdn.discordapp.com/attachments/1042...

cyberbol•8/13/25, 10:35 AM

Not sure there's much point in me posting my LoRA results anymore. Seems they serve better as comedy material, anyway

AiInfluence•8/13/25, 10:35 AM

Well i dont think he made fun of you

AiInfluence•8/13/25, 10:36 AM

rather than the "easy" available alternative

Ccyberbol Not sure there's much point in me posting my LoRA results anymore. Seems they se...

JonkoXL•8/13/25, 10:37 AM

it's a playful joke or nod on how I wanted to train a lora, and it seems a dataset just appeared

I like your lora's, and I like your posts.. I appreciate you

AiInfluence•8/13/25, 10:37 AM

i like your contribution too cyberbol

Ccyberbol Not sure there's much point in me posting my LoRA results anymore. Seems they se...

Sakami•8/13/25, 11:03 AM

You gotta have more faith in yourself – be strong

FFurkan Gözükara SECourses i am making a very big comparison

skyrrrOP•8/13/25, 11:40 AM

What's your take on buckets ? Do you notice better results without ? I use different ratios so I tend to keep them on

keyser_soze•8/13/25, 11:42 AM

anyone trained wan 2.2 in msubai trainer , particularly in h100 80gb
if you are trained on it
how can i speed up the traning time ?

FFurkan Gözükara SECourses which sampler it works best with? with IPNDM not very good

cyberbol•8/13/25, 11:48 AM

I've been testing several samplers and schedulers for the FLUX KREA model, and it's hard to choose just ONE that is the best. It really depends on your creative goal.

However, I think these combinations are definitely worth using:

Sampler: DPM++ 2M
Scheduler: bong_tangent (New)

Sampler: Euler
Scheduler: Beta / beta57

Sampler: dpmpp_2m or ipndm
Scheduler: Beta / beta57

I'm currently running a small test comparing a few options. When it's ready, I'll share the results. However, I'm also waiting for @Furkan Gözükara SECourses comparison, as it will surely be more comprehensive.

cyberbol•8/13/25, 11:51 AM

That said, it seems impossible to choose one single best combination, as it depends on many factors like the result you want to achieve, the style of the photo, and of course, personal taste. But tests like this are useful because we can rule out the samplers/schedulers that cause issues, like artifacts.

cyberbol•8/13/25, 11:57 AM

Sampler: DPM++ 2M
Scheduler: Karras

Sooo perfection, realistic... Like Me..