Software Engineering Courses (SECourses)•13mo ago

In Supir is it possible to increase VRAM usage? With my 3090 it only uses about 12GB, its tilled I t

In Supir is it possible to increase VRAM usage? With my 3090 it only uses about 12GB, its tilled I think
but I wanted to know if I can increase this so that its faster

Zhigarev•12/13/24, 12:08 PM

Tell me the most important parameters in descending order that make LoRA training more effective and faster. As I understand it, for characters, activity settings are important, while for training styles, it’s necessary for the model to memorize general style features. Accordingly, different settings for LR, Dim, and Alpha are needed—not 128 and not LR 0.0005. Or am I mistaken? Which setting greatly affects the training time? I just want to understand if, for example, image quality improves by 10-15%, but the time doubles, then I’ll use those settings

Punceka•12/13/24, 12:14 PM

Hi. For the last 2 weeks I've been training loras of people with your recommended settings. The end results was fine. This time, I'm trying to train a lora of a gas mask and no matter what I do, what recommended settings I use, it is always distorted. A lora that weights over 3gb can't replicate a simple gas mask.
For example this one here. A prompt - A man wearing a GP5M4sk
The result. Why Flux always adds a hood/a hat even if there's none in my training images or modifies the mask with objects that were trained with flux model and not mine?

TheToday•12/13/24, 2:37 PM

They have taken out to be able to train Lora in ltx video and hunyuan

dsienra•12/13/24, 3:18 PM

I tested hunyuan and is great, I'm interested on lora training is a great base model

druhl•12/13/24, 3:34 PM

I trained face on Kling and it was bad

druhl•12/13/24, 3:34 PM

Lol or perhaps my expectations are too high lol

dsienra•12/13/24, 3:51 PM

how did you trained Kling?

Ddsienra I tested hunyuan and is great, I'm interested on lora training is a great base m...

TheToday•12/13/24, 3:56 PM

https://www.reddit.com/r/StableDiffusion/comments/1hd55fx/new_lora_support_for_hunyuan_and_ltx_video/

TheToday•12/13/24, 3:57 PM

im interested too

TTheToday https://www.reddit.com/r/StableDiffusion/comments/1hd55fx/new_lora_support_for_h...

dsienra•12/13/24, 4:00 PM

Thanks, it seems is not working on windows yet

cyberbol•12/13/24, 5:46 PM

Dr Master, share with us some nice Video prompts please ?

cyberbol•12/13/24, 5:47 PM

on patreon txt file or here

�𝕏4C Hello @Furkan Gözükara SECourses i have a question regard flux lora training , w...

Furkan Gözükara SECourses•12/13/24, 7:40 PM

hi all latest updated configs are shared here : https://www.patreon.com/posts/112099700

Bbecruily In Supir is it possible to increase VRAM usage? With my 3090 it only uses about ...

Furkan Gözükara SECourses•12/13/24, 7:40 PM

nope. it uses more vram with higher resolution

Furkan Gözükara SECourses•12/13/24, 7:40 PM

what you can do is disable tiling

Furkan Gözükara SECourses•12/13/24, 7:40 PM

fp8 + disable tiling

ZZhigarev Tell me the most important parameters in descending order that make LoRA trainin...

Furkan Gözükara SECourses•12/13/24, 7:41 PM

literally i have all configs with speed

Furkan Gözükara SECourses•12/13/24, 7:41 PM

if you want more speed you can try to increase batch size

PPunceka Hi. For the last 2 weeks I've been training loras of people with your recommende...

Furkan Gözükara SECourses•12/13/24, 7:42 PM

very likely depended on your dataset or the base model flux

Furkan Gözükara SECourses•12/13/24, 7:42 PM

flux may not have sufficient good info of that in base training

Ccyberbol Dr Master, share with us some nice Video prompts please ?

Furkan Gözükara SECourses•12/13/24, 7:44 PM

i will share a technique i used - works decent but not very perfect :d

Furkan Gözükara SECourses•12/13/24, 7:45 PM

let me update patreon post for example prompts and images

Furkan Gözükara SECourses•12/13/24, 7:46 PM

check out demo v2 : https://www.patreon.com/posts/112848192

Patreon

1-Click CogVideoX1.5-5B-I2V Image to Video Installers For Windows, ...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

FFurkan Gözükara SECourses very likely depended on your dataset or the base model flux

Punceka•12/13/24, 7:56 PM

I thought if I used a unique trigger word for an object (excluding any hints it's a mask), Flux would learn it as a completely new thing and won't recognize it or try to match it to its base training images/objects. So this means that no trained object will be identical to the original one and will have some mixed Flux additions

PPunceka I thought if I used a unique trigger word for an object (excluding any hints it'...

Furkan Gözükara SECourses•12/13/24, 8:16 PM

flux has internal captioning mechanism

Furkan Gözükara SECourses•12/13/24, 8:16 PM

even if you dont provide a single caption

Furkan Gözükara SECourses•12/13/24, 8:16 PM

it still makes impact of mask word being used

FFurkan Gözükara SECourses what you can do is disable tiling

becruilyOP•12/13/24, 8:39 PM

i only tried increasing tile sizes in the tile config, how do i disable tiling in general?
do you think it will be too heavy without tiling for 3090, upscaling from 1024 pixels 4 times for example

Bbecruily i only tried increasing tile sizes in the tile config, how do i disable tiling i...

Furkan Gözükara SECourses•12/13/24, 9:52 PM

yes 4x upscale too big

Furkan Gözükara SECourses•12/13/24, 9:52 PM

you need tiled + fp8

FFurkan Gözükara SECourses hi all latest updated configs are shared here : https://www.patreon.com/posts/11...