How to Train a Highly Convincing Real-Life LoRA Model (2024 Guide)

Показать описание

This comprehensive video tutorial equips you with the knowledge to train a high-quality real-life LoRA model. Discover step-by-step techniques for breathtaking AI portraits.

🖹 Article Tutorial:

🔗 Links:

- Optimizer Extra Arguments:
- When opting for DAdaptAdam, you'll need to configure it with specific arguments: --optimizer_args "decouple=True" "weight_decay=0.01" "betas=0.9,0.999"
- For Adafactor, the recommended arguments are: --optimizer_args "relative_step=True" "scale_parameter=True" "warmup_init=True"

🕰️ Time Stamps

0:00 Intro

1:02 LoRA Training Flow

1:46 Principles Behind LoRA Training

3:46 LoRA Training Preparation

6:10 Understanding Core LoRA Training Parameters

13:36 Advanced LoRA Training Parameter Adjustments

18:44 Monitoring the Training Process

20:23 Finding the Best LoRA Model

Рекомендации по теме

Комментарии

Thanks for this great tutorial. For the current version, optimizer_args needs to be modified to this(without quotes): "decouple=True weight_decay=0.01 betas=[0.9, 0.999]". You may also need to run "pip install --force dadaptation" for DAdaptAdam if a missing module exception occurs.

sanz

This info may have been the case months ago with SD 1.5, but the newer tools, and tools for SDXL don't need cropped photos, it detect the resolution and adjusts accordingly with the training.

cmdr_talikarni

Interesting with Search & Replace plot...learned something new :-)

bornTwoFight

Thanks for pointing out details and explaining the formulas !

I already found here a detail that i was overlooking :) (Caption Extention)

TheVilkaz

Thanks for this, and I'm glad you changed to your real voice since this! :)

GoranBackmanMusic

Why did you copy another Youtuber's voice?

MariushAI

Mac, via Pinokio and Kohya v24 settings any combination which will work on macs ? i seem to have it worked out only once so far, 270 images took 3 days. Not sure if MPS not seeing as an option was the problem, for KOYHA creator suggested MPS might make it faster. sadly no one in sight to get MPS going in Kohya SS V24... for us MAC users....

Beauty.and.FashionPhotographer

крутое видео все четко и понятно молодец! щас говорят новые обучения на человека появилось...

K-A_Z_A-K_S_URALA

I don't understand the reasoning behind not using the regularization/ classification images? Is there any study that suggests otherwise? Many others are using this and producing great, realistic results. In fact, captions are harder to nail.

divye.ruhela

How does your speech synthesis process look like?

RoonyKingXL

I used the DAdaptAdam optimizer but the lora turned out black, then I made the lora with adafactor and everything was fine. Can you tell me why DAdaptAdam doesn’t work? I have rtx 4070 video card

LevAi