Software Engineering Courses (SECourses)•12mo ago

underfit means not trained enough

Furkan Gözükara SECoursesOP•1/22/25, 5:58 PM

caption probably wont help

FFurkan Gözükara SECourses underfit means not trained enough

AiInfluence•1/22/25, 5:58 PM

so more epoch or more images ? may I send you my dataset to check?

AiInfluence•1/22/25, 6:03 PM

AiInfluence•1/22/25, 6:03 PM

28 images

AAiInfluence Click to see attachment

Furkan Gözükara SECoursesOP•1/22/25, 6:35 PM

perhaps remove all tshirts

Furkan Gözükara SECoursesOP•1/22/25, 6:35 PM

but your all images are literlaly duplicate :d

AAiInfluence Click to see attachment

Timson•1/22/25, 6:35 PM

I suggest making more diverse dataset

AiInfluence•1/22/25, 6:35 PM

but this is the logo

AiInfluence•1/22/25, 6:35 PM

should I make mockups ?

Timson•1/22/25, 6:35 PM

you need 20+ images of logo in defferent contexts sizes etc

TTimson you need 20+ images of logo in defferent contexts sizes etc

Furkan Gözükara SECoursesOP•1/22/25, 6:35 PM

i agree

AiInfluence•1/22/25, 6:36 PM

should I keep ohwx logo

AiInfluence•1/22/25, 6:36 PM

as token and class ?

AiInfluence•1/22/25, 6:39 PM

https://civitai.com/articles/7486/how-to-train-a-logo-lora-for-flux

AiInfluence•1/22/25, 6:40 PM

I found this I want your opinion on it @Furkan Gözükara SECourses

AAiInfluence should I keep ohwx logo

Furkan Gözükara SECoursesOP•1/22/25, 6:49 PM

you see how it defines logo in prompt

Furkan Gözükara SECoursesOP•1/22/25, 6:49 PM

The mug is filled with a light brown liquid, likely tea, and features a stylized, hexagonal logo on its front. The logo consists of a bold, blue and white geometric design with a prominent blue 'C' in the center."

Furkan Gözükara SECoursesOP•1/22/25, 6:49 PM

that is useful

Furkan Gözükara SECoursesOP•1/22/25, 6:49 PM

during inference

AiInfluence•1/22/25, 6:50 PM

should I caption the images like him or let flux do the job?

AAiInfluence should I caption the images like him or let flux do the job?

Furkan Gözükara SECoursesOP•1/22/25, 7:04 PM

i let flux do job

FFurkan Gözükara SECourses i let flux do job

AiInfluence•1/22/25, 7:05 PM

Dr btw have you ever trained logo Lora/dreambooth?

AiInfluence•1/22/25, 7:17 PM

Btw you are right of diversity

AiInfluence•1/22/25, 7:18 PM

the last dataset which came very close to the logo was of many different people and poses wearing the shirt

AiInfluence•1/22/25, 7:18 PM

this time I made it from duplicates that's why it failed

FFurkan Gözükara SECourses Alhamdulillah first test passed and I got the screen on monitor. Time to built i...

Neuromasters•1/22/25, 7:26 PM

what's the beast?

Neuromasters•1/22/25, 7:26 PM

5090?

AAiInfluence Dr btw have you ever trained logo Lora/dreambooth?

pwndnoobcake•1/22/25, 8:05 PM

that's a field i have to explore too, so please share your insights when you are done

Ppwndnoobcake that's a field i have to explore too, so please share your insights when you are...

AiInfluence•1/22/25, 8:16 PM

diversity is what matters most

AiInfluence•1/22/25, 8:17 PM

the first model I made was pretty close and had different people wearing the shirt

AiInfluence•1/22/25, 8:17 PM

so if I added more mockups with the logo (custom made) it would be even begged

AiInfluence•1/22/25, 8:17 PM

the second model I made with duplicates was awful not recommended

AAiInfluence Dr btw have you ever trained logo Lora/dreambooth?

Furkan Gözükara SECoursesOP•1/22/25, 9:42 PM

i think not yet

NNeuromasters 5090?

Furkan Gözükara SECoursesOP•1/22/25, 9:42 PM

not yet but first day hopefully

Furkan Gözükara SECoursesOP•1/22/25, 9:42 PM

setting up the system

LMSO•1/22/25, 11:37 PM

Hi @Dr. Furkan Gözükara , I've subscribed to the patreon

LLMSO Hi @Dr. Furkan Gözükara , I've subscribed to the patreon

Furkan Gözükara SECoursesOP•1/23/25, 12:22 AM

hi welcome

Furkan Gözükara SECoursesOP•1/23/25, 12:22 AM

ty so much

Yobin•1/23/25, 4:20 AM

@Dr. Furkan Gözükara hello! In your Lora configs for flux I saw that you using float instead of fp16. Is there any issues for fp16/bf16 now?

YYobin @Dr. Furkan Gözükara hello! In your Lora configs for flux I saw that you using f...

AiInfluence•1/23/25, 8:43 AM

I had slightly less quality results with fp16, fp32 is the best

YYobin @Dr. Furkan Gözükara hello! In your Lora configs for flux I saw that you using f...

Furkan Gözükara SECoursesOP•1/23/25, 11:55 AM

hello

Furkan Gözükara SECoursesOP•1/23/25, 11:55 AM

it is save precision

Furkan Gözükara SECoursesOP•1/23/25, 11:56 AM

float is no quality loss

Furkan Gözükara SECoursesOP•1/23/25, 11:56 AM

but fp16 is almost same as well

Furkan Gözükara SECoursesOP•1/23/25, 11:56 AM

so it is not training but how you save the trained model

FFurkan Gözükara SECourses hello

Yobin•1/23/25, 12:05 PM

fp16 has smaller file size right? I just thought they're issues with fp16 like instability and what not!

Yobin•1/23/25, 12:05 PM

hence bf16 is there

pwndnoobcake•1/23/25, 2:29 PM

Hello, @Furkan Gözükara SECourses Can you please help me how to use masks in kohya? I only want to use it on a few images, is it even possible this way?

pwndnoobcake•1/23/25, 2:30 PM

i mean for training dreambooth/loras

roosterpolice•1/23/25, 2:30 PM

Hello - anyone know if there Is a reason why no choice of flux inside Run Koyha on massive compute?

underfit means not trained enough

Similar Threads

Similar Threads

Similar Threads