Software Engineering Courses (SECourses)•17mo ago

Kohya SS GUI FLUX LoRA experiments - The realism is just great and i am not even cherry picking (the

Kohya SS GUI FLUX LoRA experiments - The realism is just great and i am not even cherry picking (these are from my x/y/z grid testings for research) or best config yet (still lots of test to do and evaluate) For stylization it is still super overfit. Trying to fix that as well

Furkan Gözükara SECoursesOP•8/21/24, 8:18 PM

Rruyanisme @Dr. Furkan Gözükara, how fast is FLUX with your graphics card? Which graphics c...

Furkan Gözükara SECoursesOP•8/21/24, 8:18 PM

you need 24 gb gpu atm

Furkan Gözükara SECoursesOP•8/21/24, 8:18 PM

or cloud

Furkan Gözükara SECoursesOP•8/21/24, 8:24 PM

post updated with newest findings scheduler + sampler https://www.patreon.com/posts/110293257

Patreon

Kohya FLUX LoRA and Fine Tuning Training Full Tutorial For Local Wi...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

Furkan Gözükara SECoursesOP•8/21/24, 8:27 PM

testing flux cfg now

Ddaroth interesting, theres some weird degradation in comfyui for flux with continous ge...

J2_Nedo•8/21/24, 8:27 PM

this it happened for me too in SwarmUI, the Comfy guys are aware of this problem

Ddaroth interesting, theres some weird degradation in comfyui for flux with continous ge...

Furkan Gözükara SECoursesOP•8/21/24, 8:27 PM

wow this is a big bug

JJ2_Nedo this it happened for me too in SwarmUI, the Comfy guys are aware of this problem

Furkan Gözükara SECoursesOP•8/21/24, 8:27 PM

it is certainly a bug

J2_Nedo•8/21/24, 8:28 PM

yes sure, it happpens if you create alot of pics without stop

FFurkan Gözükara SECourses wow this is a big bug

daroth•8/21/24, 8:29 PM

I changed to a different workflow and that fixed it

Furkan Gözükara SECoursesOP•8/21/24, 8:45 PM

nice

FFurkan Gözükara SECourses Kohya SS GUI FLUX LoRA experiments - The realism is just great and i am not even...

clarkcheekykent•8/21/24, 8:46 PM

Regarding style overfitting

what did you tag your photos with? did you tag taken on smartphone or something like amateur photography?

Also have you tried adding back in generated photos of yourself (the ones with good likeness) that dont have that standard style. Or are you only experimenting with how good a model can do if you have very basic dataset

Finally another thing i found to fix style is generating first few steps with 0.4 Lora weight, then generating the remaining with 1.0 weight. Since the first few steps are dictating colors and style it gives it much better style that is closer to what original model would have produced

All the above has helped my loras be much more versatile being able to generate much more natural looking images while retaining likeness

Cclarkcheekykent Regarding style overfitting what did you tag your photos with? did you tag take...

Furkan Gözükara SECoursesOP•8/21/24, 8:48 PM

only ohwx man atrm

Furkan Gözükara SECoursesOP•8/21/24, 8:49 PM

how do you do that=? i am gonna ask swarmui :"steps with 0.4 Lora weight, then generating the remaining with 1.0 weight"

J2_Nedo•8/21/24, 9:07 PM

Dr.Furkan i can not find the menue in swarmui for the extra args settingsm any quick hint?

JJ2_Nedo Dr.Furkan i can not find the menue in swarmui for the extra args settingsm any q...

Furkan Gözükara SECoursesOP•8/21/24, 9:11 PM

display advanced

J2_Nedo•8/21/24, 9:13 PM

i meant the ExtraArgs --fast you just posted under tips-info

JJ2_Nedo i meant the ExtraArgs --fast you just posted under tips-info

Furkan Gözükara SECoursesOP•8/21/24, 9:14 PM

it is in backends

Furkan Gözükara SECoursesOP•8/21/24, 9:14 PM

Furkan Gözükara SECoursesOP•8/21/24, 9:14 PM

i think speed up fp8 significantly

J2_Nedo•8/21/24, 9:16 PM

aaah thank you!

Furkan Gözükara SECoursesOP•8/21/24, 9:16 PM

lol some configs are pure noise

Furkan Gözükara SECoursesOP•8/21/24, 9:22 PM

this one nice but resemblance reduced

FFurkan Gözükara SECourses how do you do that=? i am gonna ask swarmui :"steps with 0.4 Lora weight, then g...

clarkcheekykent•8/21/24, 10:38 PM

I'm using comfyUI and the advance sample lets you choose start and end steps, and so i have two samplers in series, one takes lower model weight and passes the output latent to the other higher model weight sampler

Cclarkcheekykent I'm using comfyUI and the advance sample lets you choose start and end steps, an...

Furkan Gözükara SECoursesOP•8/21/24, 10:56 PM

i just told this to swarmui

FFurkan Gözükara SECourses i just told this to swarmui

clarkcheekykent•8/21/24, 11:00 PM

never used swarm ui, didnt realize it can take natural language configs

Cclarkcheekykent never used swarm ui, didnt realize it can take natural language configs

Furkan Gözükara SECoursesOP•8/21/24, 11:12 PM

no i mean the developer :d

Furkan Gözükara SECoursesOP•8/21/24, 11:13 PM

out of 7 new testing 3 of them are noise

Furkan Gözükara SECoursesOP•8/21/24, 11:15 PM

just got a better config out of new tests amazing 100 epoch 1500 steps

Furkan Gözükara SECoursesOP•8/21/24, 11:17 PM

reg images killed likeliness hugely

nicearoni•8/21/24, 11:18 PM

i was noticing that as well - bottom row of Dr. Furkan's grid, 2nd from right - best, 3000 steps, with reg images - bears the worst resemblance

Furkan Gözükara SECoursesOP•8/21/24, 11:19 PM

yep

Furkan Gözükara SECoursesOP•8/21/24, 11:19 PM

i am gonna update patreon with best new configs

Furkan Gözükara SECoursesOP•8/21/24, 11:21 PM

Ddaroth I changed to a different workflow and that fixed it

J2_Nedo•8/21/24, 11:40 PM

the problem happens not only on SwarmUI even on Forge for me. Other reporting it too, like here: https://github.com/mcmonkeyprojects/SwarmUI/issues/192

GitHub

Generatiing images with LORA in Flux dev 16-bit DType renders 2nd a...

Expected Behavior Subsequent 16-bit images generated with a LORA in Flux dev should not degrade into fuzz. Actual Behavior When using any LORA, the first image generated after switching to Default ...