Software Engineering Courses (SECourses)•15mo ago

Oh so a smaller dataset will reduce vram? I tried fine tuning 128 images on a 3090 last night but co

Oh so a smaller dataset will reduce vram? I tried fine tuning 128 images on a 3090 last night but couldn't get it to come under 26.8Gb vram use which gave me an estimate of 184hrs

gave up a used mass compute as it was cheaper than my electricity at that time estimate. But it's good to know for the future to try again with a smaller dataset.

Deleted User•10/4/24, 7:57 AM

When I trained mine, it was 26 images and it still used just 23.1GB VRAM. It shouldn't be that slow for 30 images unless you have memory fallback enabled or you're going over your limit somehow.

EError-404-unknown Oh so a smaller dataset will reduce vram? I tried fine tuning 128 images on a 30...

nicearoni•10/4/24, 1:55 PM

smaller dataset will reduce overall time of training assuming you keep repeats\epochs the same, won't necessarily help with VRAM usage

joshtech•10/4/24, 3:29 PM

@Furkan Gözükara SECourses Can you recommend a NSFW capable Flux Checkpoint for training a Dreambooth Model? Or would you recommend to use the base version?

daroth•10/4/24, 4:14 PM

how would you handle captioning for Flux training, if the character has unusual features?

daroth•10/4/24, 4:15 PM

like her spots

Ddaroth how would you handle captioning for Flux training, if the character has unusual ...

crystalwizard•10/4/24, 4:41 PM

Something like "a woman with sots running down the sides of her face". The caption shoukd tell the AI exactly what its looking at so it understands what it is.

daroth•10/4/24, 4:44 PM

hm, wouldn't mentioning the spots make them a variable? so if I afterwards prompt for her with name, the markings wouldn't be there unless mentioned?

Ddaroth hm, wouldn't mentioning the spots make them a variable? so if I afterwards promp...

crystalwizard•10/4/24, 4:46 PM

Youre training it, not prompting it.

crystalwizard•10/4/24, 4:48 PM

You are creating an object that will update its weights for specific information when used. Put that specific information in you dataset lables

crystalwizard•10/4/24, 4:50 PM

And only use that object if you want it to know that specific information

Ddaroth I'm confused, how long should a full dreambooth training last (4090)? started on...

Furkan Gözükara SECourses•10/4/24, 7:07 PM

it is easy

Furkan Gözükara SECourses•10/4/24, 7:07 PM

you are very likely using shared vram

Furkan Gözükara SECourses•10/4/24, 7:07 PM

otherwise speed cant be that bad

Jjoshtech <@205854764540362752> Can you recommend a NSFW capable Flux Checkpoint for train...

Furkan Gözükara SECourses•10/4/24, 7:08 PM

i dont work with NSFW so dont know

Ddaroth like her spots

Furkan Gözükara SECourses•10/4/24, 7:08 PM

i dont think needed

Furkan Gözükara SECourses•10/4/24, 7:08 PM

i am working on a new rare token

Furkan Gözükara SECourses•10/4/24, 7:08 PM

according to t5 XXL

Furkan Gözükara SECourses•10/4/24, 7:08 PM

dumped its token

daroth•10/4/24, 9:00 PM

iiiinteresting. 25 pics is enough to train a style for Flux.

Ddaroth iiiinteresting. 25 pics is enough to train a style for Flux.

Furkan Gözükara SECourses•10/4/24, 9:17 PM

it sounds low but you can try

daroth•10/4/24, 9:20 PM

tried it via the civitai trainer, it looks promising actually. trainable artstyles with only a few dozen pic could be amazing

Ddaroth iiiinteresting. 25 pics is enough to train a style for Flux.

crystalwizard•10/4/24, 9:29 PM

there are some options that allow you to just use one image, actually

Latentnaut•10/4/24, 9:47 PM

Ey Furkan, I'm trying to Extract Flux Lora, but I only got this progress bar to 0, it's frozen
I don't know how to deal with the warning about torch, could you please help me?

DDeleted User When I trained mine, it was 26 images and it still used just 23.1GB VRAM. It sho...

Latentnaut•10/4/24, 9:50 PM

what " memory fallback enabled" is? and why is relevant to the training?

FFurkan Gözükara SECourses fine tuned on RTX A6000

Latentnaut•10/4/24, 10:01 PM

how much time did it take?

LLatentnaut Ey Furkan, I'm trying to Extract Flux Lora, but I only got this progress bar to ...

Furkan Gözükara SECourses•10/4/24, 11:55 PM

hit enter

LLatentnaut how much time did it take?

Furkan Gözükara SECourses•10/4/24, 11:56 PM

it is around 30 second per it when batch size is 7

Furkan Gözükara SECourses•10/4/24, 11:56 PM

so 100 steps would be 3000 seconds

FFurkan Gözükara SECourses i dont work with NSFW so dont know

joshtech•10/5/24, 9:45 AM

Okay, please forget about the NSFW. It is not important.

Do you think it is worth to try a community Flux Checkpoint for training a Dreambooth Model? If yes, which would you recommend ? Or would you always recommend to use the base version? Best regards and thanks a lot

dxqb•10/5/24, 10:34 AM

@Furkan Gözükara SECourses Flux anti-bleeding protection. look at right-most column. No more Furkan Obama

FFurkan Gözükara SECourses hit enter

Latentnaut•10/5/24, 11:54 AM

the issue was I put a "space" in the name of the output model, it wasn't a good idea

Ddxqb <@205854764540362752> Flux anti-bleeding protection. look at right-most column....

Latentnaut•10/5/24, 11:56 AM

I don't get it, could you tell more please? and a picture of the original woman would be great

LLatentnaut I don't get it, could you tell more please? and a picture of the original woman ...

dxqb•10/5/24, 12:52 PM

left is without lora, just Flux. the middle is regular lora training with the bleeding effect. you can see that not only the prompt "ohwx woman" was changed, but also "woman" by a lot. on the right is the lora with prior regularization: the prompt "woman" looks almost exactly like in the unchanged model with no lora on the left, only "ohwx woman" is changed

Ddxqb left is without lora, just Flux. the middle is regular lora training with the bl...

Maxivi•10/5/24, 12:55 PM

how many regularzation images did you use? did you caption your training images?

MMaxivi how many regularzation images did you use? did you caption your training images?

dxqb•10/5/24, 12:55 PM

I did not use any reg images. it's a code change of the trainer.

Maxivi•10/5/24, 12:56 PM

what code did you change?

Ddxqb I did not use any reg images. it's a code change of the trainer.

smoke•10/5/24, 1:35 PM

Very interesting since prior reg means using reg images

smoke•10/5/24, 1:35 PM

https://tenor.com/view/hmm-thinking-gif-1696455838339387033

Tenor

dxqb•10/5/24, 2:09 PM

prior just means original model. I guess the name of "prior preservation" parameter is named that because originally people used images generated by the prior model

smoke•10/5/24, 2:21 PM

So what did you change in order to avoid bleeding? Would love to experiment with this aswell

Ddxqb <@205854764540362752> Flux anti-bleeding protection. look at right-most column....

Furkan Gözükara SECourses•10/5/24, 5:28 PM

can you eloborate more? last time i tried it wasnt have any improvement :d

Ddxqb I did not use any reg images. it's a code change of the trainer.

Furkan Gözükara SECourses•10/5/24, 5:29 PM

yes i was gonna ask this

Furkan Gözükara SECourses•10/5/24, 5:29 PM

because it was not working at all last time i tested. which trainer code change fixed?

dxqb•10/5/24, 5:30 PM

OT, but it's experimental.

dxqb•10/5/24, 5:30 PM

you might be able to reach a good result with reg images too, because it does something very similar. not exactly though

FFurkan Gözükara SECourses because it was not working at all last time i tested. which trainer code change ...

crystalwizard•10/5/24, 5:32 PM

you do realize that because of what was done to flux - it really is not trainable, right?

Ssmoke So what did you change in order to avoid bleeding? Would love to experiment with...

dxqb•10/5/24, 5:33 PM

it's not a secret but this is not a development channel. I've shared on the OT discord

Ddxqb OT, but it's experimental.

Furkan Gözükara SECourses•10/5/24, 5:33 PM

so you modified OT code? can you give full details please we can forward to kohya

FFurkan Gözükara SECourses so you modified OT code? can you give full details please we can forward to kohy...

Maxivi•10/5/24, 5:33 PM

already done

Ddxqb it's not a secret but this is not a development channel. I've shared on the OT d...

Furkan Gözükara SECourses•10/5/24, 5:33 PM

can you link

Oh so a smaller dataset will reduce vram? I tried fine tuning 128 images on a 3090 last night but co

Similar Threads