Software Engineering Courses (SECourses)•15mo ago

<@205854764540362752> Flux anti-bleeding protection. look at right-most column. No more Furkan Obam

@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 ...

dxqbOP•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?

dxqbOP•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

dxqbOP•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?

dxqbOP•10/5/24, 5:30 PM

OT, but it's experimental.

dxqbOP•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...

dxqbOP•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

MMaxivi already done

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

you forwarded ? i dont see :d

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

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

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, 5:33 PM

dm on twitter i sent him

MMaxivi dm on twitter i sent him

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

he doesnt read there :d

Maxivi•10/5/24, 5:33 PM

he does

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

can you reply to this thread

Maxivi•10/5/24, 5:34 PM

he replies to me

Furkan Gözükara SECourses•10/5/24, 5:34 PM

he does but rarely

Maxivi•10/5/24, 5:34 PM

but sure

Furkan Gözükara SECourses•10/5/24, 5:34 PM

he says he doesnt get dm notifications

FFurkan Gözükara SECourses he says he doesnt get dm notifications

crystalwizard•10/5/24, 5:34 PM

from you maybe

Furkan Gözükara SECourses•10/5/24, 5:34 PM

give me related code part i will forward

Ccrystalwizard from you maybe

Furkan Gözükara SECourses•10/5/24, 5:34 PM

nope he replies back my every reply

crystalwizard•10/5/24, 5:34 PM

he obviously replies to max or max wouldn't have said he does

Furkan Gözükara SECourses•10/5/24, 5:35 PM

he also replied my dms but he said he doesnt get notifications so he replies so late

FFurkan Gözükara SECourses can you link

dxqbOP•10/5/24, 5:35 PM

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...