Software Engineering Courses (SECourses)•4mo ago

i don't know why but output character is not good it getting different character every time i used

i don't know why but output character is not good it getting different character every time i used 200 epochs and its not same character i used in dataset

JimiOP•9/2/25, 7:32 AM

i did on flux it was great result

JimiOP•9/2/25, 7:34 AM

should i try lora?

JJimi i don't know why but output character is not good it getting different character...

Furkan Gözükara SECourses•9/2/25, 8:44 AM

which base sdxl model?

Furkan Gözükara SECourses•9/2/25, 8:45 AM

by the way sdxl is no where near flux

JJimi i don't know why but output character is not good it getting different character...

JonkoXL•9/2/25, 9:22 AM

captions / learning rate?

JonkoXL•9/2/25, 9:22 AM

why are you training checkpoint on a character, that's what lora is for

JimiOP•9/2/25, 9:23 AM

i used the gonzaLomo XL model

JJonkoXL why are you training checkpoint on a character, that's what lora is for 😉

JimiOP•9/2/25, 9:24 AM

i tried lora in flux but dream boot is way better result compare that

JimiOP•9/2/25, 9:24 AM

have't tried SDXL lora yet

JonkoXL•9/2/25, 9:27 AM

Somehow I got good results with SDXL Lora trained on this (Pony) checkpoint. (It's NSFW, but seems to be giving good results)

JonkoXL•9/2/25, 9:27 AM

https://civitai.com/models/643732/animatedponyreal

JimiOP•9/2/25, 9:31 AM

do you use Kohya for SDXL lora?

JonkoXL•9/2/25, 9:31 AM

koyha_ss gui

JonkoXL•9/2/25, 9:31 AM

yes

JimiOP•9/2/25, 9:31 AM

do you have any setting to share

JonkoXL•9/2/25, 9:31 AM

those are on my old pc

JonkoXL•9/2/25, 9:31 AM

Default works I believe

JimiOP•9/2/25, 9:31 AM

ohh

JonkoXL•9/2/25, 9:31 AM

try to aim for 1600-3000 steps

JonkoXL•9/2/25, 9:31 AM

at least

JonkoXL•9/2/25, 9:32 AM

and enable sampling

JonkoXL•9/2/25, 9:32 AM

at 1000+ it should resemble whatever you wanted to train

JonkoXL•9/2/25, 9:32 AM

and at 1800 it should be decent enough

JimiOP•9/2/25, 9:32 AM

got it

JimiOP•9/2/25, 9:32 AM

will try that

JonkoXL•9/2/25, 9:32 AM

what captions do you have?

JonkoXL•9/2/25, 9:32 AM

because less is often better

JimiOP•9/2/25, 9:33 AM

i have't use any captions

JimiOP•9/2/25, 9:33 AM

should i have to use?

JonkoXL•9/2/25, 9:33 AM

try using captions

Just describe what you dont want to have in the image

JimiOP•9/2/25, 9:33 AM

okay

JonkoXL•9/2/25, 9:34 AM

so for a woman in a park with a bench and a tree, and some dog
-> myL0r4K3yw0rd, park, bench, tree, dog

JonkoXL•9/2/25, 9:34 AM

dont mention the woman or her body features , you can mention her clothes however

JimiOP•9/2/25, 9:34 AM

can i use khoya captions?

JimiOP•9/2/25, 9:34 AM

i mean blip one

JonkoXL•9/2/25, 9:35 AM

yes, but you should manually remove the ones refering to the likeness of the woman

JimiOP•9/2/25, 9:35 AM

okayy

JonkoXL•9/2/25, 9:35 AM

so " woman, young face, freckles, curvy" you should never want in the captions for example (for a character)

JimiOP•9/2/25, 9:35 AM

it will be generic

JonkoXL•9/2/25, 9:36 AM

Just describe as captions what you dont want to have in the image

JimiOP•9/2/25, 9:36 AM

ya got it

JimiOP•9/2/25, 9:36 AM

Thanks

JonkoXL•9/2/25, 9:37 AM

please share your results afterwards

JonkoXL•9/2/25, 9:37 AM

and definitely use sampling every 100 steps

JimiOP•9/2/25, 9:39 AM

okay will do that

JimiOP•9/2/25, 9:39 AM

i am trying diffrent dataset in diffrent model. i will try wan next then qwen

JonkoXL•9/2/25, 9:40 AM

wan 2.1 or 2.2?

JonkoXL•9/2/25, 9:40 AM

2.1 may be easier

JimiOP•9/2/25, 9:42 AM

2.2

i don't know why but output character is not good it getting different character every time i used

Similar Threads

Similar Threads

Similar Threads