Software Engineering Courses (SECourses)•17mo ago

OneTrainer dev is making the claim that you can have finetuning quality when using Dora-Loras, that

OneTrainer dev is making the claim that you can have finetuning quality when using Dora-Loras, that they have recently implemented and I think Kohya had for a while

dxqbOP•8/20/24, 9:34 PM

I've not tested it, just might be something interesting for you

Dazzastrous [4090]•8/20/24, 9:35 PM

Dora Loras hehe just rolls of the tongue

dxqbOP•8/20/24, 9:35 PM

it's only called DoRA, I just added LoRA to make clear that it's just another type of lora

Dazzastrous [4090]•8/20/24, 9:36 PM

All we need is Sora Dora Loras

DDazzastrous [4090]All we need is Sora Dora Loras

dxqbOP•8/20/24, 9:37 PM

paper by Liu, Wang, Yin andChen

dxqbOP•8/20/24, 9:37 PM

https://arxiv.org/abs/2402.09353

arXiv.org

DoRA: Weight-Decomposed Low-Rank Adaptation

Among the widely used parameter-efficient fine-tuning (PEFT) methods, LoRA and its variants have gained considerable popularity because of avoiding additional inference costs. However, there still often exists an accuracy gap between these methods and full fine-tuning (FT). In this work, we first introduce a novel weight decomposition analysis t...

dxqbOP•8/20/24, 9:37 PM

future of AI seems Chinese when you look at the authors of so many papers

FFurkan Gözükara SECourses swarm

cap10•8/20/24, 9:44 PM

can you use flux loras in swarm ?

DDazzastrous [4090]looks pretty good to me alrteady Doc

Furkan Gözükara SECourses•8/20/24, 10:17 PM

yep

Ddxqb OneTrainer dev is making the claim that you can have finetuning quality when usi...

Furkan Gözükara SECourses•8/20/24, 10:17 PM

dora is supposed to be

Ccap10 can you use flux loras in swarm ?

Furkan Gözükara SECourses•8/20/24, 10:17 PM

yep i am using

Furkan Gözükara SECourses•8/20/24, 10:17 PM

right now

Furkan Gözükara SECourses•8/20/24, 10:18 PM

Ddxqb https://arxiv.org/abs/2402.09353

Furkan Gözükara SECourses•8/20/24, 10:18 PM

supir is chinese model

dxqbOP•8/20/24, 10:24 PM

just installed some old hdd to have more space for .safetensors. damn this is slow. don't switch from NVMe to HDD

FFurkan Gözükara SECourses Click to see attachment

clarkcheekykent•8/20/24, 10:33 PM

It works on photo real images, because model replaces all photo real men with the trained data.

But it doesnt learn what 'ohwx' tag means. So try adding "Cartoon image of ohwx man as Aladdin" see if it still keeps some of the ohwx features

Ddxqb just installed some old hdd to have more space for .safetensors. damn this is sl...

Furkan Gözükara SECourses•8/20/24, 10:37 PM

it is painful

Cclarkcheekykent It works on photo real images, because model replaces all photo real men with th...

Furkan Gözükara SECourses•8/20/24, 10:38 PM

gonna check in a moment

Cclarkcheekykent It works on photo real images, because model replaces all photo real men with th...

Furkan Gözükara SECourses•8/20/24, 10:52 PM

looks like man prompt is strongly overwritten

Furkan Gözükara SECourses•8/20/24, 10:52 PM

probably we need reg images

Furkan Gözükara SECourses•8/20/24, 10:54 PM

ye lora is too powerful it made model overfit

Cclarkcheekykent It works on photo real images, because model replaces all photo real men with th...

Furkan Gözükara SECourses•8/20/24, 11:06 PM

really hard:D

2305-Cartoon20drawing20image20of20a20ohwx20man20in20t-flux1-dev-827753794.png

Furkan Gözükara SECourses•8/20/24, 11:10 PM

2307-Cartoon20drawing20image20of20a20ohwx20man20wear-flux1-dev-1627649612.png

Dazzastrous [4090]•8/20/24, 11:11 PM

Jeff Goldblum vibes

DDazzastrous [4090]Jeff Goldblum vibes

Furkan Gözükara SECourses•8/20/24, 11:15 PM

true

Furkan Gözükara SECourses•8/20/24, 11:15 PM

post updated : https://www.patreon.com/posts/110293257

Patreon

Kohya FLUX LoRA and Fine Tuning Training Full Tutorial For Local Wi...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

Furkan Gözükara SECourses•8/20/24, 11:16 PM

it really struggles

Furkan Gözükara SECourses•8/20/24, 11:16 PM

stylized outputs

2311-pencil20colored20Cartoon20drawing20image20of-flux1-dev-1047516507.png

Furkan Gözükara SECourses•8/20/24, 11:18 PM

time to start new 8 trainings :d

FFurkan Gözükara SECourses true

sergio_gonzo•8/20/24, 11:42 PM

Dr Furkan i hope everything is ok , STILL NOT OPTION FOR 16 gigas VRAM GPUS?

Ssergio_gonzo Dr Furkan i hope everything is ok , STILL NOT OPTION FOR 16 gigas VRAM GPUS?

Furkan Gözükara SECourses•8/20/24, 11:42 PM

yep i am messaging kohya

Ssergio_gonzo Dr Furkan i hope everything is ok , STILL NOT OPTION FOR 16 gigas VRAM GPUS?

Furkan Gözükara SECourses•8/20/24, 11:43 PM

here you can see

Furkan Gözükara SECourses•8/20/24, 11:43 PM

https://github.com/kohya-ss/sd-scripts/pull/1374#issuecomment-2299933025

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...

Furkan Gözükara SECourses•8/20/24, 11:43 PM

it uses 18 gb

Furkan Gözükara SECourses•8/20/24, 11:43 PM

but here try it

Furkan Gözükara SECourses•8/20/24, 11:43 PM

lowest_vram.json5.2KB

Furkan Gözükara SECourses•8/20/24, 11:43 PM

tell me your step speed / shared vram suage

FFurkan Gözükara SECourses really hard:D

clarkcheekykent•8/20/24, 11:48 PM

thats actually not too bad though, it atleast took some of your likeness, not sure whats going on with my snow white prompt

Cclarkcheekykent thats actually not too bad though, it atleast took some of your likeness, not su...

Furkan Gözükara SECourses•8/20/24, 11:52 PM

i am doıing more tests

Cclarkcheekykent thats actually not too bad though, it atleast took some of your likeness, not su...

�

𝔸𝖨ΞӾ•8/21/24, 12:01 AM

Have you tried setting Lora strength to like 1.4? I've noticed sometimes that's what can really force it with some loras

�𝔸𝖨ΞӾ Have you tried setting Lora strength to like 1.4? I've noticed sometimes that's ...

Furkan Gözükara SECourses•8/21/24, 12:07 AM

i am using 1.0. it is already too powerful

Furkan Gözükara SECourses•8/21/24, 12:08 AM

Furkan Gözükara SECourses•8/21/24, 12:16 AM

Raw vs eyeglasses prompt - it is interesting that FLUX is learning me without eyeglasses when not prompted. This was never case with SDXL

FFurkan Gözükara SECourses Raw vs eyeglasses prompt - it is interesting that FLUX is learning me without ey...

�

𝔸𝖨ΞӾ•8/21/24, 12:21 AM

Perhaps due to captioning with the dataset?

�𝔸𝖨ΞӾ Perhaps due to captioning with the dataset?

Furkan Gözükara SECourses•8/21/24, 12:22 AM

same i just use ohwx man

Furkan Gözükara SECourses•8/21/24, 12:22 AM

test reg images impact too

Furkan Gözükara SECourses•8/21/24, 12:24 AM

running a big test for an hard prompt

�

𝔸𝖨ΞӾ•8/21/24, 12:32 AM

oh ok yeah that makes sense now. I just figured probably one of your training images didn't have the glasses described in the caption, but then again the frames of the glasses aren't super bold and noticeable from a distance so that could be another reason

Darth Teethius•8/21/24, 12:36 AM

does anyone know if normal alternating via prompt such as [x|y] where it alternates between x and y each step is still usable in a natural language style model like flux?

sergio_gonzo•8/21/24, 12:49 AM

maybe i will wait more, so the process can make it with a gpu 16 vram, maybe onetrainer make support for flux

OneTrainer dev is making the claim that you can have finetuning quality when using Dora-Loras, that

Similar Threads