Software Engineering Courses (SECourses)•15mo ago

https://discord.com/channels/1102003518203756564/1188183242747158630/1292072916976009348 the sample

https://discord.com/channels/1102003518203756564/1188183242747158630/1292072916976009348 the sample and some discussions below that, also a code snippet

FFurkan Gözükara SECourses he also replied my dms but he said he doesnt get notifications so he replies so ...

Maxivi•10/5/24, 5:35 PM

he's also japan time zone so keep that in mind

FFurkan Gözükara SECourses he also replied my dms but he said he doesnt get notifications so he replies so ...

crystalwizard•10/5/24, 5:35 PM

maybe he jsut means he has notifications turned off then

Ddxqb https://discord.com/channels/1102003518203756564/1188183242747158630/12920729169...

Furkan Gözükara SECourses•10/5/24, 5:39 PM

ok i forwarded code to the thread

Furkan Gözükara SECourses•10/5/24, 5:39 PM

lets see now

Furkan Gözükara SECourses•10/5/24, 5:39 PM

https://github.com/kohya-ss/sd-scripts/pull/1374#issuecomment-2395126424

GitHub

support SD3 by kohya-ss · Pull Request #1374 · kohya-ss/sd-scripts

Replace SD3Tokenizer with the original CLIP-L/G/T5 tokenizers.
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...

Maxivi•10/5/24, 5:39 PM

@crystalwizard you think the same issues will happen with sd3 8b

Ddxqb https://discord.com/channels/1102003518203756564/1188183242747158630/12920729169...

Furkan Gözükara SECourses•10/5/24, 5:40 PM

did you notice any resemblance quality loss?_

Furkan Gözükara SECourses•10/5/24, 5:40 PM

also which file you added this i will test :d

MMaxivi <@407561236339752981> you think the same issues will happen with sd3 8b

crystalwizard•10/5/24, 5:40 PM

yup. it's just larger. issues don't show up quite as much with a larger model - but they don't vanish just because it's larger

Furkan Gözükara SECourses•10/5/24, 5:40 PM

if you have cloned repo i can use it

FFurkan Gözükara SECourses did you notice any resemblance quality loss?_

dxqbOP•10/5/24, 5:42 PM

not sure, it's new. just found it interesting to share that it's even possible. I am not proposing that anyone implements it at this point.

Ddxqb not sure, it's new. just found it interesting to share that it's even possible. ...

Furkan Gözükara SECourses•10/5/24, 5:42 PM

can you send me modified file

Furkan Gözükara SECourses•10/5/24, 5:42 PM

i would like to test

Furkan Gözükara SECourses•10/5/24, 5:43 PM

in onetrainer when training sdxl reg images made such a big improvement

Furkan Gözükara SECourses•10/5/24, 5:43 PM

if we get it on flux would be so amazing

FFurkan Gözükara SECourses if we get it on flux would be so amazing

crystalwizard•10/5/24, 5:43 PM

you can not train flux. you get 'something' but you don't get an object that actually does what a lora is supposed to do

crystalwizard•10/5/24, 5:44 PM

flux is broken, rigid, untrainable, and there's a reason that black forest never released training code. they KNOW it's not trainable

crystalwizard•10/5/24, 5:45 PM

what you're geting is about the same as what you'd get if you took a photo and jsut drew over it. you're not updating the photo. you're just drawing over it

Ccrystalwizard what you're geting is about the same as what you'd get if you took a photo and j...

dxqbOP•10/5/24, 5:46 PM

you couldn't be more wrong. SD15/XL is more like that. Flux is even able to learn from low res and integrate it into existing model knowledge and generate good high res

Ddxqb you couldn't be more wrong. SD15/XL is more like that. Flux is even able to lear...

crystalwizard•10/5/24, 5:46 PM

i'm not wrong. i know what the code is, i know what the neural network is. i know what they did to it to mask the issues it has, and i know it's not trainable

Furkan Gözükara SECourses•10/5/24, 5:47 PM

crystalwizard•10/5/24, 5:47 PM

and you running around creating loras are not creating loras that do anything. you're getting something that gives you image results - but you are NOT getting loras that actually work like loras are supposed to

Ddxqb you couldn't be more wrong. SD15/XL is more like that. Flux is even able to lear...

Furkan Gözükara SECourses•10/5/24, 5:47 PM

your idea looks solid from claude 3.5 :d

Furkan Gözükara SECourses•10/5/24, 5:47 PM

can you send me modified file please

dxqbOP•10/5/24, 5:48 PM

ok, but you have to be willing to experiment, I don't know any good settings yet

Ddxqb ok, but you have to be willing to experiment, I don't know any good settings yet

Furkan Gözükara SECourses•10/5/24, 5:48 PM

yep

Furkan Gözükara SECourses•10/5/24, 5:48 PM

no worries

Furkan Gözükara SECourses•10/5/24, 5:49 PM

just send me file that you got those reuslts

Furkan Gözükara SECourses•10/5/24, 5:49 PM

you shown

Furkan Gözükara SECourses•10/5/24, 5:49 PM

no need config

Furkan Gözükara SECourses•10/5/24, 5:49 PM

only modified file

Ddxqb ok, but you have to be willing to experiment, I don't know any good settings yet

Maxivi•10/5/24, 5:53 PM

https://tenor.com/view/hmm-suspect-gif-22611582

Tenor

Furkan Gözükara SECourses•10/5/24, 5:56 PM

man my massed compute 1tb not enough

daroth•10/5/24, 5:57 PM

https://tenor.com/view/what-the-hell-is-going-on-out-here-butt-head-beavis-and-butt-head-s2-e7-whats-happening-gif-11441942162150164992

Tenor

Furkan Gözükara SECourses•10/5/24, 5:58 PM

some new discovery perhaps

dxqbOP•10/5/24, 6:02 PM

@Furkan Gözükara SECourses just in case you are thinking about training TE - because that would make sense with regularization, to separate the concepts further - the code isn't ready for that. it only unhooks the lora from the transformer.

Ddxqb <@205854764540362752> just in case you are thinking about training TE - because ...

Furkan Gözükara SECourses•10/5/24, 6:02 PM

well it doesnt matter. i tested all cases all bleeds huge

Furkan Gözükara SECourses•10/5/24, 6:02 PM

and it is issue of the dit architecture

Furkan Gözükara SECourses•10/5/24, 6:02 PM

not TE or clip training

Furkan Gözükara SECourses•10/5/24, 6:02 PM

so i will train only unet to test

dxqbOP•10/5/24, 6:03 PM

what I mean is: if this reg method works well, it might even be working better with TE training. but not yet, so turn it off either way

Furkan Gözükara SECourses•10/5/24, 6:04 PM

Maxivi•10/5/24, 6:05 PM

@Furkan Gözükara SECourses seems prior preservation in kohya is broken

Maxivi•10/5/24, 6:06 PM

https://github.com/kohya-ss/sd-scripts/pull/1374#issuecomment-2395126424

GitHub

support SD3 by kohya-ss · Pull Request #1374 · kohya-ss/sd-scripts

Replace SD3Tokenizer with the original CLIP-L/G/T5 tokenizers.
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...

Maxivi•10/5/24, 6:07 PM

bghira responded

Furkan Gözükara SECourses•10/5/24, 6:08 PM

i think he didnt see my reply

becaue i added. but good

Furkan Gözükara SECourses•10/5/24, 6:08 PM

if this get fixed nice

FFurkan Gözükara SECourses kk

dxqbOP•10/5/24, 6:23 PM

ah, and batch size 1!

Ddxqb ah, and batch size 1!

Furkan Gözükara SECourses•10/5/24, 6:24 PM

batch size 1 is best

https://discord.com/channels/1102003518203756564/1188183242747158630/1292072916976009348 the sample

Similar Threads