Software Engineering Courses (SECourses)•17mo ago

yey! but speed is lower than expected. I guess 4060ti is not the best one for training

Nnovsky yey! but speed is lower than expected. I guess 4060ti is not the best one for tr...

Furkan Gözükara SECourses•8/27/24, 4:42 PM

are you using shared vram

Furkan Gözükara SECourses•8/27/24, 4:42 PM

speed will get better

Furkan Gözükara SECourses•8/27/24, 4:42 PM

wait like 150 steps to see final speed

FFurkan Gözükara SECourses are you using shared vram

novskyOP•8/27/24, 4:43 PM

I guess not

novskyOP•8/27/24, 4:44 PM

but my ram is full I guess with model and clips

Nnovsky I guess not

Furkan Gözükara SECourses•8/27/24, 4:50 PM

ye not using

Furkan Gözükara SECourses•8/27/24, 4:50 PM

wait 150 steps

Furkan Gözükara SECourses•8/27/24, 4:50 PM

and show me speed

FFurkan Gözükara SECourses and show me speed

novskyOP•8/27/24, 5:18 PM

didn't change much

novskyOP•8/27/24, 5:20 PM

@Furkan Gözükara SECourses is there any method to train lora by parts, because not always there is possibility to have pc turned on?

novskyOP•8/27/24, 5:33 PM

200 steps same

Nnovsky 200 steps same

Furkan Gözükara SECourses•8/27/24, 5:37 PM

so to speed up you have 2 options

Furkan Gözükara SECourses•8/27/24, 5:37 PM

turn off t5 text encoder masking attention

Furkan Gözükara SECourses•8/27/24, 5:37 PM

very little quality loss

Furkan Gözükara SECourses•8/27/24, 5:37 PM

speed up signficiant

Furkan Gözükara SECourses•8/27/24, 5:37 PM

if still slow reduce training resolut

Furkan Gözükara SECourses•8/27/24, 5:37 PM

like 768x768

Furkan Gözükara SECourses•8/27/24, 5:37 PM

this will hve more impact on quality

novskyOP•8/27/24, 5:38 PM

got it. Will be trying new parameters tomorrow. Will lyk results

novskyOP•8/27/24, 5:38 PM

thanks

Nnovsky thanks

Furkan Gözükara SECourses•8/27/24, 5:39 PM

you are welcome

Xcorpion•8/27/24, 8:19 PM

I was wondering, do warmup steps damage a lora?

XXcorpion I was wondering, do warmup steps damage a lora?

Furkan Gözükara SECourses•8/27/24, 8:33 PM

warmup steps means lower or higher LR

Furkan Gözükara SECourses•8/27/24, 8:33 PM

may damage

Yojimbo69•8/27/24, 9:33 PM

I am making a LoRA of a realistic character. The training data contain face swapped images and most have too much glare around the eyes. If I were to include a caption about eyes having too much glare, (say "eye glare") would the LoRA train to only keep the eye glare if "eye glare" were in the prompt, and otherwise not include the eye glare if it were not mentioned?