Software Engineering Courses (SECourses)•16mo ago

I'm just saying that even during inference a low-precision base model doesn't make a difference, so

I'm just saying that even during inference a low-precision base model doesn't make a difference, so how would it during training. The Lora weights are still bf16+

dxqbOP•9/15/24, 11:38 AM

fascinating how Flux learns without any distortion. it just slowly adjusts

Ddxqb I'm just saying that even during inference a low-precision base model doesn't ma...

Furkan Gözükara SECourses•9/15/24, 11:42 AM

it makes some

Furkan Gözükara SECourses•9/15/24, 11:42 AM

not very dramatic

dxqbOP•9/15/24, 5:50 PM

I'm getting 6s/it at home on a 4070, but 14 s/it on an A6000. Do you have any idea why this could be? identical config

dxqbOP•9/15/24, 5:50 PM

flux lora

Ddxqb I'm getting 6s/it at home on a 4070, but 14 s/it on an A6000. Do you have any id...

Furkan Gözükara SECourses•9/15/24, 6:04 PM

interesting

Ddxqb I'm just saying that even during inference a low-precision base model doesn't ma...

wardensc2•9/16/24, 12:09 AM

No, base model fp8 which uses to train lora does affect the lora. For example 2 train 2 lora same dataset but one with fp16 based model one with fp8. The lora with fp16 have better prompt understand than fp8. My training is person lora. While the person quality is still good with both loras. The prompt understanding every subjects outside of the person is reduce with fp8 lora. For example: my prompt have something like little angels playing in the background. The fp16 lora show some angel with body of human and wing but the fp8 showing some birds.

FFurkan Gözükara SECourses interesting

dxqbOP•9/16/24, 6:43 AM

was a mistake, images were cached at a different resolution on the 4070

Ddxqb was a mistake, images were cached at a different resolution on the 4070

Furkan Gözükara SECourses•9/16/24, 10:45 AM

this is actually one of my fears

Furkan Gözükara SECourses•9/16/24, 10:45 AM

different folders should be used for different resolutions

Furkan Gözükara SECourses•9/16/24, 10:45 AM

kohya may fail to properly cache

jkgnl•9/16/24, 11:18 PM

I am training currently a Lora with Flux using 20 photos and 200 Epoch, is it normal it will take 4 tot 5 hours using an RTX 4090 locally? Or did I miss any optimization setting or script?

Jjkgnl I am training currently a Lora with Flux using 20 photos and 200 Epoch, is it no...

Furkan Gözükara SECourses•9/16/24, 11:25 PM

sounds accurate

Furkan Gözükara SECourses•9/16/24, 11:25 PM

what is your step speed?

Furkan Gözükara SECourses•9/16/24, 11:25 PM

which rank config?

jkgnl•9/16/24, 11:38 PM

Step speed is 4.71s/it , how can I see my rank config?

jkgnl•9/16/24, 11:39 PM

I used a Saas solution where I trained on the same 20 photos using Flux Lora, it took me only 30 minutes, any idea what can cause the difference in speed?

Jjkgnl I used a Saas solution where I trained on the same 20 photos using Flux Lora, it...

Furkan Gözükara SECourses•9/16/24, 11:49 PM

from zip file which one you loaded?_

Jjkgnl I used a Saas solution where I trained on the same 20 photos using Flux Lora, it...

Furkan Gözükara SECourses•9/16/24, 11:49 PM

they are lower quality simple as that

Furkan Gözükara SECourses•9/16/24, 11:49 PM

unless expensive h100 one

FFurkan Gözükara SECourses from zip file which one you loaded?_

jkgnl•9/17/24, 7:09 AM

Rank_3_18950MB_9_05_Second_IT

jkgnl•9/17/24, 7:31 AM

When training instead of an person, an object or a style, would you recommend a minimal amount of images or are 20 images enough?

Jjkgnl When training instead of an person, an object or a style, would you recommend a ...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/17/24, 7:39 AM

20 should be enough

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 20 should be enough

jkgnl•9/17/24, 8:28 AM

Thanks Nitin

Jjkgnl Rank_3_18950MB_9_05_Second_IT 🙂

Furkan Gözükara SECourses•9/17/24, 10:20 AM

ye rank 3 slower

Furkan Gözükara SECourses•9/17/24, 10:20 AM

your speed good

Jjkgnl When training instead of an person, an object or a style, would you recommend a ...

Furkan Gözükara SECourses•9/17/24, 10:21 AM

for style i suggest more

FFurkan Gözükara SECourses for style i suggest more

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/17/24, 10:26 AM

I have seen people train on style using 7 images

daroth•9/17/24, 12:07 PM

I'm really happy with how good Flux can be trained, the rank 3 is perfect for the 4090, thanks Doctor

funny how despite all its capabilities, Flux if it EVER needs to show a nipple, even accidentally its so LOST that it just puts flesh colored gummy candies in their place

jkgnl•9/17/24, 12:30 PM

Dr. Furkan, could you make a video how to train multiple people, objects or styles in one training? Is it possible to train a person, a style and an object in the same time?

Jjkgnl Dr. Furkan, could you make a video how to train multiple people, objects or styl...

Furkan Gözükara SECourses•9/17/24, 1:03 PM

multiple people really hard they bleed

Furkan Gözükara SECourses•9/17/24, 1:03 PM

you need to have them in same picture

Furkan Gözükara SECourses•9/17/24, 1:03 PM

i plan that

daroth•9/17/24, 1:22 PM

we can get good people loras in around 2k steps, but what about concepts? does that require more learning? like unique physical characteristics and such

Furkan Gözükara SECourses•9/17/24, 1:24 PM

it is hard to tell that is why you should save checkpoints and compare

Furkan Gözükara SECourses•9/17/24, 1:24 PM

my 256 images was best at 80 epoch

FFurkan Gözükara SECourses multiple people really hard they bleed

jkgnl•9/17/24, 1:36 PM

Maybe I overlooked the video, do you also have a good tutorial on training with Dreambooth? When using 20 images, can I use 200 epoch like the Lora training? Is there a general rule of thumb regarding amount of pictures vs epochs?

Jjkgnl Maybe I overlooked the video, do you also have a good tutorial on training with ...

Furkan Gözükara SECourses•9/17/24, 1:36 PM

yes all same

Furkan Gözükara SECourses•9/17/24, 1:36 PM

just use my configs

FFurkan Gözükara SECourses yes all same

jkgnl•9/17/24, 1:38 PM

Based on your experience what is the recommended amount of pictures and epochs for getting the best results without overfitting? Lora/Dreambooth?

FFurkan Gözükara SECourses yes all same

jkgnl•9/17/24, 1:41 PM

In my last Lora training I used 20 pictures, 200 epochs and Rank_3_18950MB_9_05_Second_IT, it took almost 5 hours to complete. when I use 40, 60, 80 or 100 pictures, how many epochs would your recommend, is there a rule of thumb?

Jjkgnl In my last Lora training I used 20 pictures, 200 epochs and Rank_3_18950MB_9_05_...

Furkan Gözükara SECourses•9/17/24, 1:57 PM

sadly no rule of thumb

Furkan Gözükara SECourses•9/17/24, 1:57 PM

but for 15 images i get 150 epoch best

Furkan Gözükara SECourses•9/17/24, 1:57 PM

for 256 i got best at 80 epoch - it was on 8 gpu batch size 8

Furkan Gözükara SECourses•9/17/24, 1:58 PM

for 256 images 40 epoch was undertrained with batch size 1

FFurkan Gözükara SECourses sadly no rule of thumb

jkgnl•9/17/24, 2:48 PM

Thanks for your fast answer, so it's more trial and error then