Software Engineering Courses (SECourses)•2y ago

Hello @Dr. Furkan Gözükara can I please ask you 2 questions? 1. is it better to train a LORA on the

Hello @Dr. Furkan Gözükara can I please ask you 2 questions?
1. is it better to train a LORA on the base SD1.5 model instead of a realistic checkpoint? (for a realistic character)
2. can I train a LORA on SDXL if my training images are 768*768?

FFurkan Gözükara SECourses how much vram it has?

Digital [Starburst]•12/30/23, 5:08 PM

12 GB

Digital [Starburst]•12/30/23, 5:10 PM

I know you said for Dreambooth 16 GB is required, but I believe I read in the instructions that using Kohya I could get by with 12 GB (although clearly I'm pushing my luck). However, the install uses Automatic1111 and to use it I require TensorRT to run it. So I feel like a fish out of water right now.

Digital [Starburst]•12/30/23, 5:11 PM

It's running out of memory when it trys to build the engine. With each step I get the Out of Memory error and finally it crashes altogether and says Error in big red letters in the UI.

Digital [Starburst]•12/30/23, 5:12 PM

I tried building the 512 to 768 dynamic engine alone which also failed. I wanted to build the Default engine.

Digital [Starburst]•12/30/23, 5:13 PM

My CUDA and CUDNN version is 11.8. When I swapped that out for 12.3 it wouldn't work at all, so clearly that is wrong.

Digital [Starburst]•12/30/23, 5:15 PM

I'm also in the process of taking this course: https://www.udemy.com/course/master-ai-image-generation-using-stable-diffusion/ which teaches you to use SD completely from code with no UI. He claims that doing so gives you more control over certain aspects of image generation. Truthfully I think learning both aspects of SD are important. The use of UIs and the use of code because learning to code it teaches you more about how it works.

JMPA•12/30/23, 6:51 PM

Is anyone here?

JJMPA Hi guys, I put my logo on controlnet on Inpaint section, but I select de paint p...

JMPA•12/30/23, 6:51 PM

May someone answer this?

AIGambino•12/30/23, 8:09 PM

https://jsonformatter.org/4eec43 does anyone know why im getting these errors?

SHARE DATA

No Description

Kallamamran•12/30/23, 10:44 PM

@Dr. Furkan Gözükara I have 1962 reg-images i need to filter out which ones I can scale to 1024x1024. Used SOTA on the original images and these are the images it picked out from the main folder, but the images are not cropped/rescaled to 1024 :/

Kallamamran•12/30/23, 10:44 PM

Don't want to go Birme on 1.962 images

Kallamamran•12/30/23, 10:44 PM

I don't know what SOTA did because it's not all square images in the folder

Kallamamran•12/30/23, 10:45 PM

VVlad Hello @Dr. Furkan Gözükara can I please ask you 2 questions? 1. is it better to ...

Xylber•12/31/23, 12:51 AM

I'm not the doc, but I can answer the 1st one: In my experience it is better to train in the base model, and then generate images in the model you want. It not only seems to give slightly better results, but also let you change the model to generate more images without causing ugly results. But I'm sure it depends on the checkpoint you use. For example, traning in RealisticVision and then generating with AmiReal gave me worst results than training in the base model and then generating with AmiReal. My experience, of course, test yourself.

�

🕯Fox Plaid🕯•12/31/23, 1:02 AM

hello im getting this issue when trying to make a sdxl lora with my pc using the rtx 3070ti. what is this error? and what does it mean i have to solve? i followed this torturial but i cant see what im doing wrong i have followed the guide step by step.. https://youtu.be/AY6DMBCIZ3A

message.txt51.5KB

YouTubeSECourses

First Ever SDXL Training With Kohya LoRA - Stable Diffusion XL Trai...

Updated for SDXL 1.0. How to install #Kohya SS GUI trainer and do #LoRA training with Stable Diffusion XL (#SDXL) this is the video you are looking for. I have shown how to install Kohya from scratch. The best parameters to do LoRA training with SDXL. How to use Kohya SDXL LoRAs with ComfyUI. How to do checkpoint comparison with SDXL LoRAs and m...

�

🕯Fox Plaid🕯•12/31/23, 1:04 AM

this is my config

config_sdxl.json3.5KB

�

🕯Fox Plaid🕯•12/31/23, 1:13 AM

i wonder what will happen if i remove the args

�

🕯Fox Plaid🕯•12/31/23, 1:16 AM

wait it says out of memory

�

🕯Fox Plaid🕯•12/31/23, 1:16 AM

huh

�

🕯Fox Plaid🕯•12/31/23, 1:16 AM

what is going on

�

🕯Fox Plaid🕯•12/31/23, 1:18 AM

ohhhhhh it doubles up the steps

�

🕯Fox Plaid🕯•12/31/23, 1:18 AM

wth

�

🕯Fox Plaid🕯•12/31/23, 1:23 AM

man even if i try changing epoch its still out of memoery

�

🕯Fox Plaid🕯•12/31/23, 1:23 AM

this is my current config

config_sdxl.json3.45KB

�

🕯Fox Plaid🕯•12/31/23, 1:27 AM

was a resoloution issue

�

🕯Fox Plaid🕯•12/31/23, 1:31 AM

never mind 8gb vram is too small

�🕯Fox Plaid🕯never mind 8gb vram is too small

BoricuaPab•12/31/23, 1:40 AM

i was able to train my first sdxl lora yesterday using an rtx 2060 super 8gb gpu, for 1 epoch though it took 6hrs

�

🕯Fox Plaid🕯•12/31/23, 1:40 AM

Ahh okay

�

🕯Fox Plaid🕯•12/31/23, 1:40 AM

Nah i think ill wait till lower end hardware can render a lora fast

BoricuaPab•12/31/23, 1:40 AM

it does swap into shared gpu though

FiveSAI•12/31/23, 2:02 AM

Getting quite good results with sd1.5 lora training with new settings

�🕯Fox Plaid🕯Ahh okay

FiveSAI•12/31/23, 2:03 AM

you can also try training the sdxl model on 768x768

�

🕯Fox Plaid🕯•12/31/23, 2:04 AM

I did that on 768x768 but still out of memory

�

🕯Fox Plaid🕯•12/31/23, 2:04 AM

I just think the program is too overkill for my pc

AIGambino•12/31/23, 2:09 AM

Can someone tell me the setting theyre using ot train a 1.5 model I cannot get it working to save my life and nobody seems to have any insight on specific settings

�🕯Fox Plaid🕯I just think the program is too overkill for my pc

FiveSAI•12/31/23, 2:09 AM

are you using xformers and gradient checkpointing ?

�

🕯Fox Plaid🕯•12/31/23, 2:09 AM

Yes

FiveSAI•12/31/23, 2:10 AM

memory efficient attention?

AAIGambino Can someone tell me the setting theyre using ot train a 1.5 model I cannot get i...

FiveSAI•12/31/23, 2:12 AM

depends what are you aiming for ? realism?

AIGambino•12/31/23, 2:14 AM

yeah

FFiveSAI Getting quite good results with sd1.5 lora training with new settings

mikemenders•12/31/23, 7:23 AM

Can you share your training script?

Deleted User•12/31/23, 9:51 AM

--ddp_gradient_as_bucket_view

DDeleted User --ddp_gradient_as_bucket_view

Deleted User•12/31/23, 9:51 AM

Is not working in kaggle

Deleted User•12/31/23, 9:51 AM

Need too much memory

tom•12/31/23, 10:32 AM

I'm having trouble with kaggle. I tried the original Lora training video and got into the Kohya gui without any issues. But, the json file was a no go.