Software Engineering Courses (SECourses)•16mo ago

I’m training a LoRA model using Flux on pictures of a person in Kohya for 200 epochs, 1 repeat, rank

I’m training a LoRA model using Flux on pictures of a person in Kohya for 200 epochs, 1 repeat, rank 128, and batch size 1. The model turns out to be 2.5 GB, but I have a strict limit — the model needs to be under 1 GB. How can I make the model smaller so I can run it later on fal.ai? I understand that I can lower the rank, but what other approaches are there?

EErik I’m training a LoRA model using Flux on pictures of a person in Kohya for 200 ep...

Furkan Gözükara SECourses•9/19/24, 1:46 PM

save as fp16

Furkan Gözükara SECourses•9/19/24, 1:46 PM

then you can reduce network dimension

Furkan Gözükara SECourses•9/19/24, 1:46 PM

also you can try lora dynamic resize - may work better than training 64 rank

Furkan Gözükara SECourses•9/19/24, 1:46 PM

but i havent tested dynamic resize yet

ErikOP•9/19/24, 1:47 PM

Thank you sir!

xlads•9/19/24, 2:34 PM

https://www.instagram.com/reel/DAGPTsZMLlA/?igsh=MzRlODBiNWFlZA==

evolving.ai

This is wild

Kling 1.5 just dropped, with a new feature called Motion Brush. The Motion Brush tool works by allowing users to add controlled movement to their video generations, similarly to what Runway released in 2023.

This new AI-powered feature works by allowing users to “select” an area or subject, then choose a direction, and then Kli...

Likes

7877

Instagram

Furkan Gözükara SECourses•9/19/24, 2:35 PM

kling next level

Furkan Gözükara SECourses•9/19/24, 2:52 PM

i am gonna add single layer training configs

Furkan Gözükara SECourses•9/19/24, 2:53 PM

checking vram right now

uisato•9/19/24, 3:37 PM

https://www.youtube.com/shorts/W6xmpX-fdmg

YouTubeuisato

Audioreactive Video Playhead + Realtime MIDI Control - #touchdesign...

Audioreactive Video Playhead system, now with real-time MIDI control + 21GB of new timelapses, and SD configurations.

LK + UBridge + Smartphone → TDAbleton → TouchDesigner

You can access these project files, plus many more systems, tutorials, and experiments, through: https://linktr.ee/uisato

Alittle Snack•9/19/24, 3:47 PM

Would love an opinion from the great minds here - want to train a Flux LoRA with 2 characters (want them in 1 LoRA so I can generate with them both in the same image) Thoughts on the better approach - no captioning (like what was mentioned in the 256 image tests)and train the txtenc or fully caption the images and skip the txtenc? (PS: really appreciate all the info and help thus far!)

FFurkan Gözükara SECourses fine tuning / dreambooth best

liorftw•9/19/24, 3:56 PM

Hey you've said you training a person concept with 20500 images, by that you mean you are doing fine tuning?

AAlittle Snack Would love an opinion from the great minds here - want to train a Flux LoRA with...

Furkan Gözükara SECourses•9/19/24, 4:10 PM

you need to have them in same image

Furkan Gözükara SECourses•9/19/24, 4:10 PM

otherwise they get mixed / bleed

Lliorftw Hey you've said you training a person concept with 20500 images, by that you mea...

Furkan Gözükara SECourses•9/19/24, 4:11 PM

yes it is general fine tuning with our reg images dataset

FFurkan Gözükara SECourses otherwise they get mixed / bleed

Alittle Snack•9/19/24, 4:12 PM

I can do that - but would you fully caption or just the minimal like you did int he 256 image set?

AAlittle Snack I can do that - but would you fully caption or just the minimal like you did int...

Furkan Gözükara SECourses•9/19/24, 4:12 PM

captioning really doesnt help much

Furkan Gözükara SECourses•9/19/24, 4:12 PM

since it internally captions

Furkan Gözükara SECourses•9/19/24, 4:13 PM

for training a specific concept i dont suggest caption

FFurkan Gözükara SECourses for training a specific concept i dont suggest caption

Alittle Snack•9/19/24, 4:13 PM

thanks!

hish•9/19/24, 5:08 PM

Hello,
I have a question
I'm using Stability Matrix.
I installed Stable Diffusion Forge;
and it installed automatically ControlNet, and it's built in, called ControlNet Integrated.

The issue is, I want to use "t2iadapterstyle_sd14v1" model but, under there is no "Clip Vision" under the "Preprocessors" menu

__
when I run the Stable Diffusion Forge webui, I see a "ControlNet Integrated" extension, but under the Proprocessor menu, I don't see many preprocessors, like "Hed", "Clipvision" etc..
instead I see other proprocessors like "softedge_pidinet", CLIP-G (Revision), depth_anything_v2 and many many more

Hhish Hello, I have a question I'm using Stability Matrix. I installed Stable Diffusio...

Furkan Gözükara SECourses•9/19/24, 5:27 PM

i dont use Stability Matrix it could be reason

umuthasanoglu•9/19/24, 6:26 PM

I get this error when I press start training on Kohya Flux Lora training:

umuthasanoglu•9/19/24, 6:26 PM

File "<frozen runpy>", line 198, in _run_module_as_main
File "<frozen runpy>", line 88, in _run_code
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Scripts\accelerate.EXE__main.py", line 4, in <module>
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Lib\site-packages\accelerate\commands\accelerate_cli.py", line 19, in <module>
from accelerate.commands.estimate import estimate_command_parser
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Lib\site-packages\accelerate\commands\estimate.py", line 34, in <module>
import timm
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Lib\site-packages\timm__init.py", line 2, in <module>
from .models import create_model, list_models, is_model, list_modules, model_entrypoint, \
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Lib\site-packages\timm\models__init__.py", line 28, in <module>
from .maxxvit import *
File "C:\Kohya_GUI_Flux_Installer_v35\kohya_ss\venv\Lib\site-packages\timm\models\maxxvit.py", line 225, in <module>
@dataclass
^^^^^^^^^
File "C:\Python312\Lib\dataclasses.py", line 1275, in dataclass
return wrap(cls)
^^^^^^^^^
File "C:\Python312\Lib\dataclasses.py", line 1265, in wrap
return _process_class(cls, init, repr, eq, order, unsafe_hash,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\Lib\dataclasses.py", line 994, in _process_class
cls_fields.append(_get_field(cls, name, type, kw_only))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\Lib\dataclasses.py", line 852, in _get_field
raise ValueError(f'mutable default {type(f.default)} for field '
ValueError: mutable default <class 'timm.models.maxxvit.MaxxVitConvCfg'> for field conv_cfg is not allowed: use default_factory
21:24:53-982797 INFO Training has ended.

message.txt2.37KB

Uumuthasanoglu I get this error when I press start training on Kohya Flux Lora training:

Furkan Gözükara SECourses•9/19/24, 6:52 PM

Python312

Furkan Gözükara SECourses•9/19/24, 7:37 PM

new feature coming

Furkan Gözükara SECourses•9/19/24, 7:37 PM

Tacite•9/19/24, 8:26 PM

So glad to be a patreon subscriber when i see all these good tools

Furkan Gözükara SECourses•9/19/24, 8:31 PM

@dx

Furkan Gözükara SECourses•9/19/24, 8:31 PM

@dxqbYD you were right about alpha

Furkan Gözükara SECourses•9/19/24, 8:31 PM

we need to keep it same

Furkan Gözükara SECourses•9/19/24, 8:31 PM

still 16 rank performs way worse than 128 rank in terms of resemblance

Furkan Gözükara SECourses•9/19/24, 8:31 PM

but it is lesser overfit

Furkan Gözükara SECourses•9/19/24, 8:32 PM

for styylized images generation can be useful

Marrvelous•9/19/24, 8:40 PM

Is there a tutorial to combine two different LORA models together that have different trigger words. For example, an ohwx model of myself and a bbuk model of my wife. I'd like to combine them to make pictures including both of us. When I select both LORA's while generating a picture, each with a weight of 1, I mostly get the person from the first LORA selected